Hi HN. I wrote this post after getting frustrated by the lack of ways to run the new Gemma 4 Drafter models, and mainstream tools not prioritizing this, and hiding all the performance levers.
I ended up getting a modern 26B MoE model (Gemma 4) running at reading speed on an old recycled server with a single Xeon E5-2620 v4 and 128GB of DDR3 RAM (and no GPU). It took a lot of work, but it actually worked out somehow.
I've also linked the quants at the end, but they're not gonna run unless you use the ik_llama-cpp fork I mention, see other posts for more details.
I'm not an ML engineer, so I'm by no means an expert, and the server is busy acting as a Nix cache, but if you have any question, I can try to answer, but best effort.
"-t 8 matches physical cores. The machine has 16 SMT threads but only 8 cores. On a memory-bound workload, oversubscribing threads adds scheduling cost without adding throughput: the cores are waiting on DDR3, not on each other."
But ... isnt that a classic use case for SMT? Giving T1 sth. to do while T0 is waiting on DDR(3) and vise-versa?
I also dont understand the explanation of "--cpu-moe".
If an expert has ~ 4.0 GiB of Parameters, why does optimizing the sequence of experts minimize cash trashing? With 20 MiB of L3 Cash vs 4.0 GiB of Parameters, it wont cash any noticeable amount of the Parameters, will it?
As mentioned by others, only some Intel Xeon E5-2xxx v4 did support DDR3, and according to Intel, the E5-2620 v4 is not one of them.
> But ... isnt that a classic use case for SMT? Giving T1 sth. to do while T0 is waiting on DDR(3) and vise-versa?
Waiting in terms of latency. When the bus is mostly empty and it takes a while to make a round trip it's great to try to find a few extra passengers to put on it. When the buses are all completely full adding the extra riders just makes the bus stop that much more chaotic.
This is ironically a pretty solid use case for (ex VLIW research) ILP-optimizing compilers.
Given knowable runtime hardware usage patterns (huge bursts of memory bandwidth saturation) and a single limited core/thread-shared resource (memory bandwidth), one could optimize for the constraint ahead of runtime.
Because most of the performance optimization levers you have available to pull are (a) trade compute for memory bandwidth (e.g. compression), (b) preload when memory bandwidth is available, (c) optimize the choice of what's in cache when, (d) align to cache size / memory boundaries.
Or tl;dr, try to approximate GPU ISAs at the CPU compiler level. (Which why would anyone but hobbyists, because everyone else just buys pallets of Nvidia/AMD or designs their own ML chips?)
I won't speak for cafkafk, but I have two E5 (v3/v4) systems one on DDR4 and one on DDR3. This generation of CPU all support DDR4, but a few skus do support DDR3 also. ChatGPT told me they were niche products to meet specific customer needs.
I just picked up the DDR3 board, an Aliexpress "XD3" so I could reuse some DDR3 ram on a better CPU. Quad channel 1866MT/s is not bad!
One thing to note: These Xeons have quad memory channels, that usually means double the bandwidth of an equivalent desktop CPU, if you populate all the slots.
I have a dual E5-2667 v2 server with 512GB DDR3 and it's quite nice, the memory bandwidth is higher than of a DDR4 desktop with a way newer CPU, even though it's ECC and registered.
Noted, and agree (it looks like it has also already been clicked, which I dislike). I honestly I need to redo the themes.
> You say it runs "at reading speed". Have you benchmarked it?
At some point a few weeks ago, yes I think so, but I didn't write it down for some reason... so I'll have to find a time when it's not busy and do it again without a noisy system. Right now the system is noisy, but that said doing it like this:
llama_print_timings: load time = 83911.65 ms
llama_print_timings: sample time = 26.99 ms / 128 runs ( 0.21 ms per token, 4742.15 tokens per second)
llama_print_timings: prompt eval time = 343.41 ms / 7 tokens ( 49.06 ms per token, 20.38 tokens per second)
llama_print_timings: eval time = 10639.36 ms / 127 runs ( 83.77 ms per token, 11.94 tokens per second)
llama_print_timings: total time = 11114.98 ms / 134 tokens
So 11.94 tokens per second while it's also playing binary cache and CI builder.
When I do it properly, I'll add it to the blog as well!
And if you ever run out of things to do in your copious free time, it looks like that PR #1744 was merged without the has_target_ctx assert two days after you uploaded your drafter quants. So you can now redo all your quants and rerun all your benchmarks ;-).
llama-bench is part of the llama-cpp package, but from recent experimentation, the settings it is able to (or is documented to?) accept lag behind somewhat. Not sure whether it would accept all of the esoteric settings in the article?
I'm pretty sure eval time is token generation time where it's actually outputting new tokens. If you're getting a thousand per second on that, I'd love to know on what.
From the prompt timings above, it seems like 'prompt eval time' is the equivalent to 'processing time for input tokens'.
Hyperscalers can perform this evaluation very quickly because evaluation can be significantly parallelized. The layer `i` output of token `j` only requires access to the layer `i-1` output of all previous tokens, so a parallel frontier develops. Token (0,0) [(token, layer)] is processed first, then tokens (0,1) and (1,0) can be processed in parallel, then (0,2), (1,1), and (2,0), and so on.
The maximum parallel width becomes equal to the number of layers in the model. Gemma 4 26B-A4B model discussed in this article evidently has 30 layers, giving a 30-fold speedup if the system were otherwise unconstrained (all layers can be run in parallel, and one full set of layer outputs is completed in the KV pass for each pass of the parallel sweep).
In the specific output above, however, the input prompt is only seven tokens long so there are probably considerable non-amortized spinup effects at play.
Seven tokens long input isn't very realistic, is it? For coding tasks it's normal for the input to be thousands or 10s of thousands. If it wasn't for prefix caching it'd be one miserable experience, but even then at the very best the input is often in hundreds each time. And don't even try to dump some logs into the prompt.
> Seven tokens long input isn't very realistic, is it?
The test prompt above was "Why is the sky blue?", so there's the seven tokens. I meant to highlight that because I'd expect processing of a thousand-token input to be faster per token than presented.
Something doesn't add up here. As someone who has only recently built a home-server from an E5-26xx v2 on DDR3 RAM (because I have a sh*tload of 32g DDR3 DIMMs), I can confidently say that the newer cores (E5-26xx v3 and v4) only run on DDR4 memory...
So either you have a v2 instead of a v4 (and run on DDR3 memory), or you have a v4 but with DDR4 memory (not DDR3)
There are some OEM-only v3/v4 parts with dual memory controllers (because of a RAM supply crunch at the time, funnily enough), but the E5-2620 v4 is not one of them. The classic example is the very popular 12-core E5-2678 v3.
It looks like Supermicro had some DDR3 Xeon v3/v4 boards, and the first thing that came to mind was a Shenzen workstation/gaming board using recycled parts... haven't searched on that but it's bound to exist.
> So either you have a v2 instead of a v4 (and run on DDR3 memory), or you have a v4 but with DDR4 memory (not DDR3)
Yup that's odd... I've got a Xeon 2680 v4 (14 cores) (amazing bargain of a little beast btw) and it's indeed on DDR4 and I saw all Xeons v4 as supporting DDR4 only.
Full spec (brand/model/mobo type) would have been nice: mine's an HP Z440 workstation repurposed as a server (which I only turn on when I'm working and which I religiously turn off before going to bed).
How many watts is that setup? Cool you got it to work, but maybe only useful for vintage / retro computing rather than practical if the energy consumption makes it economically wasteful.
IDK about OPs setup, but I run a pile of E5-2683v4 Xeon recycled servers for Ceph and self hosted business SaaS usage.
One node's ipmitool sensor report (and self-monitoring PSU, so grain of salt, but my UPS side monitoring tracks closely), reports 250-300w average power use. This though, mind you is for running 22 spinning disks, 2 SAS/SATA SSDs, and 4 NVME ssds, and 768GB of DDR4.
Mid-gen 2015ish Xeons were not great at power reduction, but if you are pegging the cores, they were never particularly slow, and they did have lots of PCIe lanes. This boils down to the CPU/mobo itself not being that big a cost floor, especially if you have high utilization rates.
As a comparison, my main desktop development machine, running a Threadripper 9970X, 128GB of DDR5, a RDNA4 GPU, and a small pile of NVME drives has a power floor of roughly 250W. Some CPU centric workloads you'll definitely lose out on on the older gens of machines, but they are by no means impractical.
Maybe for a desktop usecase they are absolutely suboptimal nowadays, but for a lot of realworld usecases I would say they're still relevant.
---
Like the author posts for the LLM usecase, I think optimizing the hardware choice to the application and not leaving levers unpulled is a big key, especially considering how wide a variety of bandwidth/power draw/peak frequency/corecount SKUs exist in the Xeon lines. Without knowing what you intend to run and fitting the correct processor to it, you will end up with a disappointingly poor environment fit.
How many kWh to fabricate a brand new machine better suited to the task?
As long as performance is useable (apply your own metrics!), pulling it from existing hardware is likely the option with the lower eco footprint.
Also: chances are it'll only be used for this purpose occasionally, and/or for a short while. In that scenario [fabricating new hardware] always has the bigger eco footprint.
I don’t know why you’d assume that an older system is lower footprint.
If you’ve got something consuming 100 watts average over your 24 hour period, and your electricity costs 20 cents per kWh, you’re already spending almost as much as a Claude subscription.
Just on electricity, this assumes your hardware never fails and you never incur any additional costs.
There’s a big reason why newer more efficient hardware is in demand. Something that’s 10+ years old has drastically worse performance per watt.
Obviously I am not saying to throw away your old hardware as a rule but there is a point where some of this old stuff just isn’t even worth running.
I have two LARGE Xeon systems of this era that I used to use when I was heavily involved with Kubernetes and needed to build out a home lab. One is 2x Xeon w/ 256 GB of ram, and one is 1x Xeon w/ 512GB of ram. Both are slow as dogs, and both of them take up at least 150+ watts with only one power supply. My 12th gen Intel Nuc is so, so much faster and efficient. I'm recycling the Xeon systems.
Xeon is a group of products with really varying specs. There is no indication of which XEONs. Also new consumer CPUs often have really small internal caches.
Would you consider improving the website's layout? Right now I find it below average quality and very distracting. Whether you are an engineer or not is not really important; great engineers can write horrible text or use a layout that is not ideal, for instance.
We’re not there yet, but the obvious endgame of the present bubble insanity is open models running on local hardware and devices are “good enough” for most use cases. That will completely implode what’s going on at the moment in tech.
Happened to me. CoPilot changing prices prompted me to cancel my CoPilot subscription and install a local coding model running entirely in VRAM. Will call Claude APIs when I get really stuck, but I should be able to handle 80% of my needs with a dumber local model.
For a long time, too. Programming languages rarely change much, techniques rarely change, so I should be able to use said model for I hope at least five years; and if at any time they optimize local models to cram even more intelligence into the same amount of VRAM, I can upgrade to that.
> Will call Claude APIs when I get really stuck, but I should be able to handle 80% of my needs with a dumber local model.
I experiment with all of the local models I can fit into 32GB of VRAM and I have subscriptions to multiple SOTA providers.
The difference between them is very large, unfortunately. The local models can handle small tasks and refactoring mostly okay, but doing anything challenging with them becomes a waste of time. Unfortunately the waste isn’t immediately obvious because they will come back with something that looks like it works, but then on closer examination I need to throw it out and reset them in a usable direction.
This. OpenAI and Anthropic are ultimately compute infrastructure plays and not really AI. Everyone will have models, they'll have the ability to run them. This is why the GPU shortage is in their favor.
And like Google and Meta, these companies are going to morph into advertising giants. Advertising is an economic black hole and it eats everything that comes close.
Embedding ads in LLM responses is something researchers are having a lot of trouble figuring out right now.
I have seen the results of some early attempts. It fails in such hilarious ways that all these companies are scared of productizing it. But once someone does it, the taboo is broken and everyone else will follow suit immediately.
How does that view align with Anthropic leasing data centers from others?
I don’t know OpenAI’s infra, but to the extent they are buying GPUs and building data centers with their own money, that sounds like a bad move.
Satya has mismanaged the AI transition in many ways, but one thing he got right is that models are commodities, and the value is in applications that apply them to create user benefit. I agree that any company trying to build a moat with a model is not long for this world.
Do you think there will still be an incentive to release weights in that scenario? Everyone will have models only if there continue to be companies releasing weights.
Companies won't but I suspect this is a role that something else open source-y will fill that niche. Maybe orgs like wikimedia or internet archive, maybe some hackers just making things, maybe nation states that want to disrupt other players. Also model training will get better and better both on the algo and the hardware side. You can easily see a world where you might be able to train a good enough model on a home lab in a few days.
But you will need training data. Like a whole Internet search engine or massive data scraping. That‘s a thing that will not change with better algorithms, hardware or cheaper energy.
Data is the only moat but they'll be starting in the same place the current set of players statyed out just a few years ago. I suspect that the delta between what is publicly available (if not legally publicly available! see scihub) and what open ai and anthropic have is relatively small.
Maybe. But if we can all run our own model locally in 2 years on commodity hardware OpenAI and Anthropic will start to look like WeWork during the pandemic
I agree with you that they are headed in that direction! The GPU shortage is (I think) similar to the pandemic era hiring binge. It's less about the extra compute and more about denying the GPUs to potential competitors. They're racing against time to find something that gives them real moat (gen ai I guess?) and they are trading money for time.
This is also why the money being poured into datacenters isn't going to result in as much development as you think. It's about leveraging other people's money to lockdown more future hardware. This is going to end exactly like fiber build out in the 2000s. Eventually that fiber got used but the folks who originally paid for it got hosed.
If you mean releasing model weights: They won't, because they know the "shill something" vector will get abliterated immediately. And they can't use trade secrets or copyright to stop it, either, because they released the model themselves and you don't need to redistribute weights, just an adblocker LoRA.
You just described the absolute nightmare scenario for the newly minted trillion-dollar companies whose only hope is for enterprises and SMB to move all their business processes to the cloud, with employees competing at token maxxing.
I wouldn't say "completely implode", too much money was poured int it, but it's clear we're heading in that direction. You get a model that is "good enough", plus privacy, plus savings in the long term.
Paradoxically, the better results we get from general harness of coding agents, the less moat Claude and co. get. It's unbelievably how fast some open models outpaced frontier models of just a few months ago.
I run my word processing software on my apple 2 (a total joke of a computer) instead of running it on the WANG.
I run my book keeping software on visicalc instead of the IBM.
I run my simulation software on my IBM PC (I even paid for the 8087!) instead of the VAX.
Moore's law has, at least so far, allowed the pioneers with toy computers to grow their toys big enough to solve "big boy" problems after some time has allowed the toy computers to be faster and the pioneers have scaled their crappy home-grown solution to solve their 60% of the problem that was originally solved by some enormous complex system.
Eventually the toy infrastructure gets expensive and solves 90-120% of the "big iron" problem space, but it also grows to cost as much as the big iron solution, but then a new generation of toy software and toy systems emerges to disrupt the "big iron" systems.
You're right Moore's law has been holding up, but will hit a hard limit on process node size, so all scaling will be based on multiple cores. OTH, computing per watt spent has been plateauing. If the future bottlenecks are energy and cooling, that will require infrastructure-scale solutions. My bet is this is going to be real AI company moat.
Under appreciated requirement for this to work in post-cloud times: open source
If a vendor can SaaS a solution, then enterprise is generally happy (they don't want to have to hire folks for maintenance), and that completely locks out any ability to run locally.
Between enterprise's ambivalence and the obvious financial incentive to vendors, you get SaaS-only products.
It's a huge difference. If you had AI sufficiently good running locally on a phone, you could devise workflows for things like basic digital hygiene, technical assistance, and tedious tasks like inbox management, image sorting, device updates, and so on. Privacy and security gets a big boost past some local competence threshold, and we're nearly there.
Make the local AI competent enough to do good image generation and editing, realtime voice and music generation, handle agentic tasks with a framework like Hermes, and you can take your AI places to do tasks in contexts that are inaccessible to or inappropriate for cloud.
Frontier big platform models will be the best, but there's a level of "good enough" for local uses that we're already seeing flourish, and "good enough" for the average joe is almost here.
Phones and laptops are terrible devices for local AI, way too constrained by bad thermals and small batteries. MiniPC's (many of them using mobile hardware) don't have that particular issue, and can easily run on a 24/7 basis.
The same devices that do what would be considered absolutely monstrous amounts of image and other signal processing just a few years ago. And have an always-on radio.
That level of local AI is also more or less what you need for competent autonomous robots, too. If your household robots are orchestrated from your phone, the local security and cloud convenience converge on a single device. No extra servers, etc, reduced cost, all that - local AI is a massive market amplifier.
It's a little different because cloud and blogs didn't actively get in the way of your home compute. To wit, the various cost spikes for hardware.
People -- WANT -- this technology on their home devices and (apparently?) the providers of this tech don't seem to be running a profit so they probably don't want the maintenance tail on their side either.
I think it's a bit different. Inevitable that this becomes a household-run thing? Not likely.
Running an LLM locally is theoretically viable. Running your blog on your laptop is never viable (unless you hook it up like a server). One just requires compute while the other a stable network.
The primary feature of a blog or any website is that it is available around the clock, that is the primary feature of cloud: around on the clock computer and network that scales on demand.
The primary feature of "AI" is to process information and reason with a natural language interface at speed, the primary feature of AI bigboys is to provide the machinery that runs the "models".
I disagree. We are currently in a weird period where these frontier AI companies are losing tons of money even on the subscription-based AI models. It's just too compute intensive and there's no way most people are going to be buying the kind of hardware required to run $20 worth of inference every day.
Sadly - it's going to be ads. Advertising is going to get in there and enshittify the whole thing because as always, advertising income is too easy and too plentiful for any company to resist.
Right now the models are fairly agnostic, but we are a hair-breadth away from ChatGPT responding with, "the right tool for this job is a circular saw - something like the Milwaulkee M18, which happens to be on sale at Home Depot this weekend."
$20/day x 250 days per year x # devs/agents/etc = $$$. About $5k per dev at that daily use case.
Enough to validate repurposing an existing workstation with enough RAM, or finding a used high VRAM GPU, or in my case buying a Strix Halo system for home lab and local models.
The future is once again not cloud based, for AI tools.
Most people are running a whole lot less than $20's worth of tokens per day on cloud platforms. (Is that assuming a frontier model? 1M output tokens per day?) Local hardware could easily take up that workload, at least the part of it that's non-time-critical.
The advertising future looks like that to me, too. Service proxies like OpenRouter might talk about price optimization, maybe some ad filtering. But I expect proxies will have malicious entries, too, surreptitiously altering agentic prompts.
Ads are usually the workaround where you don’t deliver enough value to get people to subscribe or payments are unavailable for some reason.
It makes sense to show some ads and get some money at low volume (like a faraway reader wanting to read a story in your local newspaper) but taking money from regular users directly will pay much more.
Newspapers are happy to cannibalize 99% of their ad revenue with a paywall if that 1% subscribes because that’s how much more money you make from someone paying $10-$20/month vs ads.
But yeah, if people use it as a buying recommendation engine, that’s where the money is on ads/referrals but a lot of AI use has little/no connection to buying intent touchpoints.
Newspapers had no choice after craigslist and later Google/Facebook took all their classified revenue.
LLMs may or may not be able to cover their costs with it. We'll see - I suspect product placement as recommendations will become a thing as it won't take as much GPU to give a "recommendation" on "the best widget for X". I firmly expect it to become enshittified the same way google and amazon search has.
I find that hard to believe. The AI companies will want to control what's possible and find new things to do that "need" their services. Otherwise it would be like Intel and Microsoft had decided in the year 2000 that computers are "good enough" now and we would have explored what's possible with that hardware ever since.
> Otherwise it would be like Intel and Microsoft had decided in the year 2000 that computers are "good enough" now and we would have explored what's possible with that hardware ever since.
I think you've misunderstood what good enough means in the context - which is a model capable of completing the tasks assigned to it without having the breadth of full generalization. Your analogy breaks down because of this - we did get 'good enough' spec profiles for different hardware. That thing you're wearing on your wrist won't have the same specifications as the box you use to play games.
I think you've misunderstood the analogy. Just ignore it, analogies mostly break down anyways.
> a model capable of completing the tasks assigned to it
The thing is, the "task assigned to it" is changing with improved capabilities. If everyone around you in 2036 is using general AI to do amazing stuff, you will probably have little interest in vibe coding slop like it's 2026.
> The AI companies will want to control what's possible and find new things to do that "need" their services.
That's correct. The problem is they have smart people, tons of money, and several years to figure that out, and the best thing they can come up is a coding agent.
That isn’t the best thing they’ve come up with. It’s a marquee product that is fit for public consumption, however.
The ‘best’ things are;
- fuzzy pattern matching algorithms for traffic analysis, human and other image target recognition.
- targeting algorithms that identify ‘suspicious’ individuals in large volumes of metadata.
- fraud analysis
- antagonistic image and video generation, both for fooling other fraud analysis, but also for propaganda, screwing with other actors, etc.
- directed high speed content generation (text, pictures, video) to spam the ‘algorithm’ and allow near realtime identification of additional buttons to push for given target audiences.
- massive marketing/ad manipulation.
Those budget line items (and the suppliers) really want to stay off the radar however, as it makes their life harder.
>Otherwise it would be like Intel and Microsoft had decided in the year 2000 that computers are "good enough" now and we would have explored what's possible with that hardware ever since.
That would be the dream... no fucking Electron! No lockdown modules.
Gamers Nexus has a good video on this, but if NVIDIA exits the consumer market, and honestly why would they stay when they can charge up to a 100x for the same wafer space for enterprise, AMD would likely do the same.
Only Apple really makes consumer hardware suitable for running things locally then, and maybe some weird Qualcomm ARM chip for Windows.
It will be hard running things locally if nobody is supplying the hardware.
Nice post and technically impressive work. I agree we need to understand the build pipeline and be able to do things locally. However, depending on your electricity cost, it might not make sense financially. These old servers are not energy efficient at all (I'm guessing that old Xeon server will easily pull 200W on load), and that model is currently at 0.1$/0.3$ per 1M tokens (with 76 tps and 262k context) in Openrouter (also, these servers are LOUD).
EDIT: I stand corrected, 200W is apparently way too high of an estimate. I used to run a bunch of old Xeon servers and they slurped watts like crazy, but I can't remember which ones exactly those were.
OK, then you're in luck. I had a bunch of old 1U rack servers and even in the next room it was too annoying to run them (they had a bunch of 40mm fans which always ran at full speed, because in a server room, no one can hear you scream).
Yeah, 1u is gonna do that. Get something that can accommodate a big tower air cooler such as the Hyper 212 and your airflow will be quieter than the disks.
I don't run it anymore but my old server was a dual xeon (with two of those coolers crammed in) and I rarely heard a peep out of it.
Could it just be really bad cooling? Looking at 9800X3D, it seems like it's running in a similar range wrt TDP unless you really push the 9800X3D. I'm comparing with desktop cpu's because that's what my workload is. cpu governor is set to performance (no schedutil). No audible change in fan speed during heavy compilation or gaming (very silent humming), and i don't have any fans beside cheap intake, cpu and exhaust fans (1 each) + an excessive amount of dust.
These servers had no fan control whatsoever, they always ran full blast. That's not untypical for rack servers, because as written: they are designed for server rooms, and you're supposed to wear ear protection there anyway... Yes, I could've modified them, but I ditched them because running them simply made no sense (especially the high idle power consumption was ridiculous).
Only when you remove it from the original server or enable low fan mode (if available). Most 1U/2U cases will happily blow at full speed well over 90db.
You likely need to replace the flow-through server chassis system with an active "normal" cooler to achieve a bit of silence.
85W might be about right. My old server CPU is in the same ballpark and compiling kernels it reached about 90w in power usage. If you want to keep it running: idle is not very low power unless you have one of the "low power" L versions, keep that in mind.
Get a 4U case, many options if you want to combine it with a NAS. Not hard to cool and keep somewhat quiet. If you can store it in a closet or something that helps too.
Well, you can use it for lots of other things as well.
Compared to the cloud you can probably save up to buy a new server every month. And don't underestimate the gains of having something to experiment on and play with.
These servers are loud if you're trying to fit them into a 1U or 2U, which requires high speed fans to generate the necessary static pressure to push air through the case. I run a similar setup in a 4U case with slow 120mm fans and it's fine.
Glad to see other people realizing this. I've been running Gemma 26B-A4B Q4 on a 2012 Xeon with 16GB to 24GB of RAM in a container. It's getting around 8 to 12 tokens per second. Obviously it's not comparable to huge contexts and running it on a GPU and the image decoder in llama.cpp is super slow compared to a GPU but for some small automation tasks and general trivia questions it's decent. The speed is just enough to not have to wait for it to finish so you can read along.
Here's my setup. You may want to figure out what the best optimizations are for your specific CPU like AVX2 because mine didn't have most of them. I did try MTP briefly but I wasn't getting performance improvements. You could play around with the batch sizes for cache or context or go even lower for Q2 and don't overcommit on threads either, but I would suggest either defaults or trying out llama-bench. This isn't by any means the best I assume but it worked decently for me and I sometimes swap out Gemma for Qwen. You could also lower q8_0 to q4_0 for more context but it could hurt quality some say, altough I have noticed it too on some models.
I'm setting up a Frankenstein system at the moment. It's a Chinese DDR3 X99 motherboard with a 12 core Xeon v3, 32gb 1866MT/s ram, and a 1080 Ti.
I'm shoehorning it back in the Optiplex that donated the ram, so it's not ready to go at the moment, but when I had it running on top of the motherboard box as a test I ran the (9B?) gemma4:e4b-it-q4_K_M since it can fit entirely in the 11gb vram. It flew, more than 50tk/s. A model that small isn't useful for coding, but there could be uses. I'd love to figure out a Wake-on-Use and use it as my personal ChatGPT. I'm not sure how that would work... Maybe proxy the LLM thru a Pi with a script to Wake-on-LAN the PC? It'll be a fun weekend project someday.
My always-on LLM is the dense Gemma4:31b that's not quite half in GPU on a 12gb 2060. It's really slow, but the quality is great and my use case is an automated queue so I'm not sitting there watching the output. I have another 2060 but unfortunately the PC won't POST with both installed for some reason.
Result is ~12 tokens per second, as reported by OP down in these comments here.
An impressive effort, and better than I would have thought possible on this hardware -- but still pretty far short of what one needs for an satisfactory interactive session.
Especially if you consider those smaller models are really cheap and fast on platforms like openrouter. Often by the factor 100-500 cheaper than SOTA models, and 2-5x in TPS.
What intrigues me the most about AI progress, is not AGI or the model du jour by $AI_UNICORN, but rather what can be run locally. I remember having an amusing, but rather useless model in a beefy gaming PC that I had 6 years ago; and now, something that’s a hundred times better on my M5 laptop.
Should the market react to the memory shortage, the progress of the Apple silicon continue at the same pace, and what we’ll be able to run locally in 6 years will be very exciting. or frightening.
Also I don’t know what this means for the valuation of the AI companies. I remember asking about this very idea to one of their employees at an event and instead of answering he bailed out to grab a cocktail.
- There is no "moat" (lasting, easy-to-defend technological edge) in AI model businesses. There are just short-term advantages.
- An AI business is a capital-intensive business, just like old factories. Data centers are expensive, models are energy-hungry, and the hardware inside must be replaced every 3–4 years.
- Smaller, specialized models eat margins from below. Transcription, voice, or image detection do not need large models.
There is no reason to expect high margins like you can in traditional software business. Benefits of AI go mostly to consumers.
edit: There is potential for economies of scale. Few megacorps can strive for cost advantage when they achieve scale (Microsoft, Google, Amazon and Meta)
It does seem like the structural characteristics we’ve observed so far suggest there is a kind of flywheel from short-term to long-term advantage due to the capital requirements at various levels.
If you’re Nvidia, making the best GPUs today, the expanding wavefront of demand is consuming them with volume and margins to give you a huge edge in building out the best next generation of GPUs. Similar to how the mobile wave gave TSMC sustained advantage for about a decade now.
I’m guessing this is also what we’re seeing as Anthropic and OpenAI swap spots in the token-vendor market.
I can see the fly wheel in action for Nvidia[1], but in terms of model building - I think the companies that have the advantage here are not Anthropic or OpenAI, but rather companies with substantial revenues from other sources - Google is the obvious player here - reported to be planning on spending 185 billion this year without having a raise a dime from the markets, but there are plenty of other companies - like Meta or Alibaba who can easily fund the longer game from existing revenues.
What you can run locally in consumer hardware is progressing pretty well.
If you get a not-quite-the-best gaming GPU like a 5080, you can run local models that are better than the state of the art from early 2025. Depending on what you want to do, you might have to switch models. The one size fits all huge models are still a data center thing.
Its a convenience thing. You can run a whole lot of stuff locally from wikipedia to social media/email/video servers whatever. Most people with a full time job and 2 kids dont do it cause who has time and energy to patch and maintain the ever growing complexity of this stuff. These systems will keep growing complex. That also means more bugs. Age old tradeoff between freedom and convenience.
You can run mediawiki at home but you won't have wikipedia. You can run a video server but you won't have all the movies that Netfix has. A local model is actually the real thing.
--what this means for the valuation of the AI companies
Probably nothing. Most users have no idea what an LLM is or how it runs. Anecdotally speaking, I see many LLM users default to whatever their day job provides to them. And even slightly more sophisticated users seem ok with paying for their openai or anthropic subscriptions.
Maybe we will see a small but dedicated group of open weight model users who prefer local llm, but everybody else will just consume from the big providers? The scenario might look something like OS choices today - a small, committed group of Linux users vs the vast majority of other users running Windows, MacOS, or Chrome?
This has always been true of software, particularly games. You can get a 5-6 year old game for a fraction of the price, and run it on modest hardware. But the industry wont sit on its hands for 5 years, there will be newer software that requires better hardware.
A new game is a totally new world with everything created from scratch. A creation. A model, on the other hand, is a reinterpretation machine for hundreds of years of human creations, but not a creation in itself, more like a discovery.
You would think that by now we would have a much better Bitcoin that's taking over the payment networks of the world but what we actually got is a shitload of shitcoin.
The E5-2620 v4 is great. Have been using it for 10 years now. Wanted to upgrade until I saw current prices. I have 64 GB ddr4. Paired it with rx 9060 xt 16 GB and games run as fast as ever. Perhaps the cpu is a slight bottleneck in DOOM The Dark Ages, but i'm at 60 fps, so no problem. Light llm on the gpu is a nobrainer, and it's cool to see that things can be tuned to run ok on the cpu. I bought 2667 v4 a month ago for 30$. I'd expect it to give a decent performance boost but I just haven't had the need for it yet, but pushing into llm like in the article I'd probably upgrade because 2667 can handle slightly faster ram.
> The E5-2620 v4 is great. Have been using it for 10 years now.
10 years? Damn, that is a long time. I always assumed that heat-induced damage will kill a CPU after a certain amount of time (5-7 years). Am I wrong here? I assume yes. Or are CPUs must stronger/tougher than the bad old days?
This is among the "real" differences between workstation/server CPUs and commodity chips for laptops/desktops/handhelds.
Even then, if a commodity chip isn't pushed full tilt at all times, and assuming that the venting and dissipation are adequate, a commodity chip can last a long time.
A quick search on Xeon production yields that it goes through a rather rigorous testing. I wouldn't be surprised that server cpu's in a desktop pc works longer. I can't overclock it either, and that probably helps with its lifespan as well. But yeah, the fact that it actually powers on when i click the button and isn't a limiting factor after 10 years is quite something.
Back from my old overclocking days - its heat that kills life. And if you keep that under control (what ages is the heatpaste, replace it ever so often) i very much doubt you'll have any life issues from the cpu itself.
Bearings in fans, caps etc. are also stuff that you need to keep an eye on.
I just replaced a i5-660 thats been powered on since 2010 24/7, heatpaste was fucked so it crashed during heavy loads :)
I wish this were somehow tagged with AI, so I would know that it's not about say, general computing or cost-efficiency (e.g. using an old xeon machine from ebay instead of new, in these cost-conscious times.)
As it is, the title is click-bait for me, as 1) it says I need at least a Xeon somehow and 2) as it doesn't say what I actually need it for.
I want to share something strange. I found a typo or two in the post and this absolutely delighted me, because it implies a human wrote the words. (Or was at least heavily involved in the editing.)
I felt like I had lost something valuable when I switched to mostly AI based programming, because I used to make so many mistakes that the computer would often do truly magical things I did not even realize were possible.
e.g. one time I tried making a collaborative drawing application but I messed up the logic, and the brush strokes would just get temporarily mirrored between the client and server, so you'd see it getting drawn over and over again in a loop.
The drawing wasn't stored anywhere, it existed only in the network packets between client and server. Accidental GNU.
AIs already make typos, not directly intentionally. Since they are token-based, and tokens are lexemes, they can misconjugate works or make grammatical errors.
When comparing hardware, the output of these tools is very helpful to let others put it into context. The post says the output is "reading speed" but knowing the prefill and token generation speeds would be a lot more helpful.
I've got an old HP Z-620 workstation with dual E5-2697 v2 CPUs (24 cores total, 48 threads @ 2.7GHz) and 128GB of DDR3 RAM. The docs say it supports up to 192GB, but I wasn't able to get it to POST with all the RAM slots full.
It's still a "homelab" beast and does great with development and GIS/Mapping applications. I was not able to figure out how to run AI workloads on it with decent performance, however, so I finally broke down and got a dedicated GPU for it. It's pretty great what can still be done with older hardware.
Did some try to estimates what it would take to bake interference for a capable large language model into silicon so that one can pipeline inputs through it and produce outputs at one token per clock cycle?
No RAM. Instead of having a general purpose multiplier that multiplies an input with a weight stored in RAM, just have a multiplier that hardcodes the weight. In some sense replace each weight with a specialized multiplier and wire them together with accumulators and activation functions in between. And some registers for pipelining. If one goes for four bit quantization, one could have sixteen optimized multipliers, one for each possible weight, and the one just selects and connects them according to the model weights and structure.
Example. If you have a neuron with 16 inputs each 8 bit wide and with a 4 bit weight per input, you will have 16 specialized multipliers each scaling its input by the corresponding weight and then the 16 scaled inputs feed into an adder tree and finally an activation function.
The memory controller is integrated into the CPU, so the motherboard chipset is irrelevant. There are some OEM-only v3/v4 parts with dual memory controllers, but the E5-2620 v4 is not one of them.
Old hardware is surprisingly effective. I've been considering a side hustle selling offline AI to local businesses who are privacy-sensitive. Medical, legal, places like that.
At the low end, I'd use old Xeons with gobs of DDR3, install some V100s, run a smaller agent for general chat inquiries, and a frontier model for the deeper stuff, with a router that passes between them depending on the complexity.
The frontier model would perform very slowly, but if it's a deep task the user can submit it in a batch in the evening e.g. "Correlate all of these cases and look for patterns" then receive the output with morning coffee.
Of course, AI helped me work out a plan for this. Haha
I've been running various models on a Mac Pro 2013 (8 cores, 32 GB RAM) at about 8 to 10 t/s for months. It's not fast, but it's more than enough for many actual tasks, in particular background tasks. An iMac pro will do just as well I suppose.
The sort of task you don't expect to end immediately. If extracting data from a bunch of PDFs takes 1 hour or the whole night, that doesn't make much difference to me.
It's not fast enough for auto completion and slightly too slow for chat (but bearable IMO).
Doesn't accepting 100% of the MTP draft tokens mean you should just be using the smaller model? Usually the acceptance rate in Qwen36 at least is around 60-70% and the "wrong" tokens are still filled in entirely by the base model, but when you just accept 100% of the draft tokens it seems kind of self defeating unless I'm wrong.
Also I feel like everyone leaves off prompt processing/prefill speeds in these articles. If you are using a very small prompt and asking for mostly generated tokens, sure but I'd love to know the time-to-response of asking for an analysis of an image or a few hundred lines of code.
As far as I know, speculative decoding still verifies that the proposed tokens are what the "big" model would generate, it just uses the guesses to make that process faster. Setting the probability threshold too low then shouldn't affect correctness, just speed (time will be wasted verifying bad guesses).
None of those settings set the speculative decoder to accept 100% of drafted token. I assume you are looking at --draft-p-min 0.0, if so, you are misunderstanding what it does.
It depends on the type of MTP. If you're using two models, draft + full, then arguably yes, the larger model isn't providing much benefit if you really are seeing 100% acceptance rates. There are other forms of speculative decoding that work within the larger model by itself though, eg. Qwen has additional speculative decoding attention heads, so there is no secondary drafting model.
Does this mean my 15 year old Phenom is too old? But it has 16 gb of DDR3 RAM!
Admittedly web browsers and it don't get along that well. Literally the only thing that drags though on my Slackware 15 system, and even then usually only when it gets to around 15 or so open tabs.
Very intriguing. This might be the use for my e5-2430 V2 X2 server that's been lying around. DDR3 is (relatively) cheap now too. Could fit 192GB of RAM in it and play around for much cheaper than a new GPU.
Honestly, at this point you're probably looking at a smaller model, for the Gemma series I'd go with Gemma 4 E4B with drafters, but that's just a hunch from using it on my laptop (where I do have a RTX 4060 M and 96gb ram).
So you'd change the invocation slightly here, but a lot of things you can potentially reuse.
That said, the Gemma 4 E4B models have so far in my experience been... not great when it comes to long context, but they are very passable for basic tasks, and even seem surprisingly okay at tool calls.
Have you tested Qwen3.6 35B? Putting aside the capability claims for that model (which I support, but are not my point here), that 35B has smaller active parameter count than the gemma 4 26B, potentially making both prefill and decode faster out of the box, and has MTP heads built in the model and well supported (you may need to make sure you download a quant that didn't strip them off, as some do to preserve space). I would be curious to see your numbers there too. And if you do test this, please go for a clean one and not a fine-tuned one.
i tried the Q4_K_M model form unsloth with your Q4_K_M drafter, but the required memory to load everything is 72GB. odd. otoh i could load Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled.IQ4_XS.gguf and it requires just ~18 GB:
llama_print_timings: sample time = 45.28 ms / 404 runs ( 0.11 ms per token, 8921.67 tokens per second)
llama_print_timings: prompt eval time = 949.42 ms / 51 tokens ( 18.62 ms per token, 53.72 tokens per second)
llama_print_timings: eval time = 24067.08 ms / 400 runs ( 60.17 ms per token, 16.62 tokens per second)
llama_print_timings: total time = 242192.55 ms / 451 tokens
so i wonder why the params used by the quantified qwen model use way less memory than the ones of gemma.
I have run llama.cpp on an i7-2600 with a 1050. It's too slow for everyday usage but it's not too slow to make it obvious AI is going to be everywhere and in everything. It's too easy to run.
Loading will take some minutes, but at 96 you can squeeze the model in and have some headroom around like ~10 GB, although depending on the Xeon, you may have to downgrade to E4B instead. Should still work thou.
I think one overlooked advantage of older Xeon systems is their availability. Many people can experiment with local AI deployments at a fraction of the cost of building a brand-new setup.
That is a very fair point! I just ran a not very scientific benchmark with the system under load, and posted the raw logs in a sibling comment above, but the short answer is that it's hitting 11.94 tokens per second for generation - while it's also being a binary cache and CI build server.
Totally just vibes based, I think it goes up to 20+ tps when it's not under load (and that's me trying to be conservative). For context, reading speed at 250 wpm would be around 5 to 6 tokens per second.
There are ways to trade off compute power for memory bandwidth (like MTP and other speculative decoding approaches). The CPU and GPU would need to be able to share the same cache for this to work. In the Strix Halo case the GPU has a private cache on the GPU die I think, which is the snag.
If you get the inference engine to route the heavy matrix math to the GPU and the speculative drafting to the CPU without choking on latency it's probably gonna be very fast.
Would love to see the benchmarks if someone actually pulls something like that off.
As someone doing this for fun on a windows 11 machine (96gb ram, 5090 24gb) I wonder if I need any flags to keep the model in memory and avoid swapping to ssd?
I use LM studio and qwen3.5 35B - but never figured out if it is swapping or not.
Om am unrelated note, does anyone know a model that can help with this use case:
Well, lets get started. I have 4 of those machines, and they are Two dual processor. They all had 32GB of ram, so now I have two with 64GB, and two with zero. They all hand stock K5000s, now how two have two cards. I stripped the uni processors ram and video cards, and put those into the dual procs. They have 256Gb SSDs, and two 1TB disk drives. One machine has 8Gb of VRam across two cards. Dual processors are 8Cx2 and 32 Threads. They can easily play 16 videos at once. For AI, I have not found a model that I can get above 3 tokens a second. Not a one.
When you use page up and page down key when reading that blog the first line on the screen is obscured by the floating bar or what ever it is. It is not even needed for reading.
The webpage's layout is just horrible. Scrolling is also
non-default - and thus rather annoying; I had to stop after
two scroll events. Why do people think they need so much
fancy effects or non-standard behaviour, if their alleged
goal is to get information across to other people?
I appreciate the downvotes without any reasoning. It's a fact that newer Intel CPUs have Intel ME which was not in older CPUs and significantly increases attack surface if you are not living in a five eyes state.
In a server, you have to worry about the ME only if you also have an Intel Ethernet interface, which is connected to a potentially hostile network.
If that is not true, the ME cannot be controlled remotely.
The existence of the ME is much more worrisome in laptops, where the ME can be accessed remotely through WiFi. There, to be certain that there is no way for the ME to be accessed remotely you would have to disconnect or cut the internal antennas and use a USB dongle for WiFi.
As five eyes citizen you have at least some rights on paper and you can appeal to your government, but if you are foreigner these guys can go gloves off without any fear of retribution.
Try analyzing Epstein files and posting about it, they'll give you a proper penetration test of all your devices to see what you found out about their ex employee.
Nowadays even EU citizens migrating away from US cloud providers are a "national security issue".
You can run these on a turing machine. At what point is it not worth it? At some point the energy to generate each token matters. We often seen token per second. I think a missing metric is tokens per kilowatt. That is what really matters.
> The argument for speculative decoding is stronger on CPU than on GPU.
Uh. Uuuh.
No?
___
Also
> While a GPU has a massive pool of ultra-fast High-Bandwidth Memory (HBM), a CPU relies on small, lightning-fast “caches” (L1, L2, L3) built directly onto the processor chip.
What purpose does the quoting of "caches" serve there?
Is this AI writing written by that model running on that host?
Hi HN. I wrote this post after getting frustrated by the lack of ways to run the new Gemma 4 Drafter models, and mainstream tools not prioritizing this, and hiding all the performance levers.
I ended up getting a modern 26B MoE model (Gemma 4) running at reading speed on an old recycled server with a single Xeon E5-2620 v4 and 128GB of DDR3 RAM (and no GPU). It took a lot of work, but it actually worked out somehow.
I've also linked the quants at the end, but they're not gonna run unless you use the ik_llama-cpp fork I mention, see other posts for more details.
I'm not an ML engineer, so I'm by no means an expert, and the server is busy acting as a Nix cache, but if you have any question, I can try to answer, but best effort.
"-t 8 matches physical cores. The machine has 16 SMT threads but only 8 cores. On a memory-bound workload, oversubscribing threads adds scheduling cost without adding throughput: the cores are waiting on DDR3, not on each other."
But ... isnt that a classic use case for SMT? Giving T1 sth. to do while T0 is waiting on DDR(3) and vise-versa?
I also dont understand the explanation of "--cpu-moe". If an expert has ~ 4.0 GiB of Parameters, why does optimizing the sequence of experts minimize cash trashing? With 20 MiB of L3 Cash vs 4.0 GiB of Parameters, it wont cash any noticeable amount of the Parameters, will it?
As mentioned by others, only some Intel Xeon E5-2xxx v4 did support DDR3, and according to Intel, the E5-2620 v4 is not one of them.
> But ... isnt that a classic use case for SMT? Giving T1 sth. to do while T0 is waiting on DDR(3) and vise-versa?
Waiting in terms of latency. When the bus is mostly empty and it takes a while to make a round trip it's great to try to find a few extra passengers to put on it. When the buses are all completely full adding the extra riders just makes the bus stop that much more chaotic.
This is ironically a pretty solid use case for (ex VLIW research) ILP-optimizing compilers.
Given knowable runtime hardware usage patterns (huge bursts of memory bandwidth saturation) and a single limited core/thread-shared resource (memory bandwidth), one could optimize for the constraint ahead of runtime.
Because most of the performance optimization levers you have available to pull are (a) trade compute for memory bandwidth (e.g. compression), (b) preload when memory bandwidth is available, (c) optimize the choice of what's in cache when, (d) align to cache size / memory boundaries.
Or tl;dr, try to approximate GPU ISAs at the CPU compiler level. (Which why would anyone but hobbyists, because everyone else just buys pallets of Nvidia/AMD or designs their own ML chips?)
Fantastic practical achievement!
I wonder if I could get similar or even better performance from similar Dell T7610 workstation with dual Xeons and also 128GB DDR3?
The CPUs are better core wise, but that probably does not make much difference?
It has CPUs 2 × Xeon E5-2697 v2
Cores / threads 24 cores / 48 threads total
Per-CPU cores 12 cores / 24 threads
Base clock 2.70 GHz
Max turbo 3.50 GHz
It is sitting gather dust but reading spead Gemma sounds promising.
You sure you got DDR3 .. I have 2 e5 v4 rigs at home and both have ddr4 ... Unless I am wrong and 2011-3 supports ddr3 and ddr4
I won't speak for cafkafk, but I have two E5 (v3/v4) systems one on DDR4 and one on DDR3. This generation of CPU all support DDR4, but a few skus do support DDR3 also. ChatGPT told me they were niche products to meet specific customer needs.
I just picked up the DDR3 board, an Aliexpress "XD3" so I could reuse some DDR3 ram on a better CPU. Quad channel 1866MT/s is not bad!
The first two generations supported DDR3 only. Haswell and Broadwell (v4) brought DDR4 support.
right, and they talk about "v4" which is DDR4.
This seems remarkably suited to my situation,
Also with 128G. Does 8 dimm sockets imply more actual bandwidth in practice?This poor thing is currently a YouTube watching box.
One thing to note: These Xeons have quad memory channels, that usually means double the bandwidth of an equivalent desktop CPU, if you populate all the slots.
I have a dual E5-2667 v2 server with 512GB DDR3 and it's quite nice, the memory bandwidth is higher than of a DDR4 desktop with a way newer CPU, even though it's ECC and registered.
(purple on black is really hard to read)
You say it runs "at reading speed". Have you benchmarked it?
> (purple on black is really hard to read)
Noted, and agree (it looks like it has also already been clicked, which I dislike). I honestly I need to redo the themes.
> You say it runs "at reading speed". Have you benchmarked it?
At some point a few weeks ago, yes I think so, but I didn't write it down for some reason... so I'll have to find a time when it's not busy and do it again without a noisy system. Right now the system is noisy, but that said doing it like this:
llama-cli --model gemma-4-26B-A4B-it-Q8_0.gguf --model-draft gemma-4-26B-A4B-t-assistant-GGUF/wikitext-2-raw_ik-llama-mtp_drafter-conservative/gemma-4-26B-A4B-it-assistant-Q8_0.gguf --spec-type mtp --draft-max 3 --draft-p-min 0.0 --color -sm graph -smgs -sas -mea 256 --split-mode-f32 --temp 0.7 --cpu-moe -t 8 --flash-attn on --mla-use 3 --merge-up-gate-experts --special --mlock --run-time-repack --spec-autotune --no-kv-offload --parallel 8 --jinja -p "Why is the sky blue?" -n 128
Gives:
So 11.94 tokens per second while it's also playing binary cache and CI builder.When I do it properly, I'll add it to the blog as well!
And if you ever run out of things to do in your copious free time, it looks like that PR #1744 was merged without the has_target_ctx assert two days after you uploaded your drafter quants. So you can now redo all your quants and rerun all your benchmarks ;-).
> two days after you uploaded your drafter quants. So you can now redo all your quants and rerun all your benchmarks ;-)
2010s Javascript, putting down the controller: Ha, no one will ever surpass my high score for wasting programmer time with dependency churn...
2026 Open Source ML: Hold my beer.
What's time to first token? Raw throughput is usually not the problem in local setups in my experience.
I am pretty sure llamacpp have their own benchmarking binary that you can use.
llama-bench is part of the llama-cpp package, but from recent experimentation, the settings it is able to (or is documented to?) accept lag behind somewhat. Not sure whether it would accept all of the esoteric settings in the article?
20 tokens per second for eval time is the killer here. It means you can't use this to process any meaningful amount of text.
A GPU typically processes close to 1000 tokens/s during eval.
The prompt is literally "why is the sky blue?" and consists of 7 tokens.
It's probably too small for the timings to be taken seriously.
I'm pretty sure eval time is token generation time where it's actually outputting new tokens. If you're getting a thousand per second on that, I'd love to know on what.
From the prompt timings above, it seems like 'prompt eval time' is the equivalent to 'processing time for input tokens'.
Hyperscalers can perform this evaluation very quickly because evaluation can be significantly parallelized. The layer `i` output of token `j` only requires access to the layer `i-1` output of all previous tokens, so a parallel frontier develops. Token (0,0) [(token, layer)] is processed first, then tokens (0,1) and (1,0) can be processed in parallel, then (0,2), (1,1), and (2,0), and so on.
The maximum parallel width becomes equal to the number of layers in the model. Gemma 4 26B-A4B model discussed in this article evidently has 30 layers, giving a 30-fold speedup if the system were otherwise unconstrained (all layers can be run in parallel, and one full set of layer outputs is completed in the KV pass for each pass of the parallel sweep).
In the specific output above, however, the input prompt is only seven tokens long so there are probably considerable non-amortized spinup effects at play.
Seven tokens long input isn't very realistic, is it? For coding tasks it's normal for the input to be thousands or 10s of thousands. If it wasn't for prefix caching it'd be one miserable experience, but even then at the very best the input is often in hundreds each time. And don't even try to dump some logs into the prompt.
> Seven tokens long input isn't very realistic, is it?
The test prompt above was "Why is the sky blue?", so there's the seven tokens. I meant to highlight that because I'd expect processing of a thousand-token input to be faster per token than presented.
I meant prompt eval time.
Something doesn't add up here. As someone who has only recently built a home-server from an E5-26xx v2 on DDR3 RAM (because I have a sh*tload of 32g DDR3 DIMMs), I can confidently say that the newer cores (E5-26xx v3 and v4) only run on DDR4 memory...
So either you have a v2 instead of a v4 (and run on DDR3 memory), or you have a v4 but with DDR4 memory (not DDR3)
Everything else doesn't work
There are some OEM-only v3/v4 parts with dual memory controllers (because of a RAM supply crunch at the time, funnily enough), but the E5-2620 v4 is not one of them. The classic example is the very popular 12-core E5-2678 v3.
This is not true. A few well known brands made both DDR3 and DDR4 servers that support v3 & v4 chips. Ask me how I know :-)
enlighten us
It looks like Supermicro had some DDR3 Xeon v3/v4 boards, and the first thing that came to mind was a Shenzen workstation/gaming board using recycled parts... haven't searched on that but it's bound to exist.
Yeah, the Intel reference page only lists DDR4, not DDR3:
https://www.intel.com/content/www/us/en/products/sku/92986/i...
> So either you have a v2 instead of a v4 (and run on DDR3 memory), or you have a v4 but with DDR4 memory (not DDR3)
Yup that's odd... I've got a Xeon 2680 v4 (14 cores) (amazing bargain of a little beast btw) and it's indeed on DDR4 and I saw all Xeons v4 as supporting DDR4 only.
Full spec (brand/model/mobo type) would have been nice: mine's an HP Z440 workstation repurposed as a server (which I only turn on when I'm working and which I religiously turn off before going to bed).
How many watts is that setup? Cool you got it to work, but maybe only useful for vintage / retro computing rather than practical if the energy consumption makes it economically wasteful.
IDK about OPs setup, but I run a pile of E5-2683v4 Xeon recycled servers for Ceph and self hosted business SaaS usage.
One node's ipmitool sensor report (and self-monitoring PSU, so grain of salt, but my UPS side monitoring tracks closely), reports 250-300w average power use. This though, mind you is for running 22 spinning disks, 2 SAS/SATA SSDs, and 4 NVME ssds, and 768GB of DDR4.
Mid-gen 2015ish Xeons were not great at power reduction, but if you are pegging the cores, they were never particularly slow, and they did have lots of PCIe lanes. This boils down to the CPU/mobo itself not being that big a cost floor, especially if you have high utilization rates.
As a comparison, my main desktop development machine, running a Threadripper 9970X, 128GB of DDR5, a RDNA4 GPU, and a small pile of NVME drives has a power floor of roughly 250W. Some CPU centric workloads you'll definitely lose out on on the older gens of machines, but they are by no means impractical.
Maybe for a desktop usecase they are absolutely suboptimal nowadays, but for a lot of realworld usecases I would say they're still relevant.
---
Like the author posts for the LLM usecase, I think optimizing the hardware choice to the application and not leaving levers unpulled is a big key, especially considering how wide a variety of bandwidth/power draw/peak frequency/corecount SKUs exist in the Xeon lines. Without knowing what you intend to run and fitting the correct processor to it, you will end up with a disappointingly poor environment fit.
How many kWh to fabricate a brand new machine better suited to the task?
As long as performance is useable (apply your own metrics!), pulling it from existing hardware is likely the option with the lower eco footprint.
Also: chances are it'll only be used for this purpose occasionally, and/or for a short while. In that scenario [fabricating new hardware] always has the bigger eco footprint.
I don’t know why you’d assume that an older system is lower footprint.
If you’ve got something consuming 100 watts average over your 24 hour period, and your electricity costs 20 cents per kWh, you’re already spending almost as much as a Claude subscription.
Just on electricity, this assumes your hardware never fails and you never incur any additional costs.
There’s a big reason why newer more efficient hardware is in demand. Something that’s 10+ years old has drastically worse performance per watt.
Obviously I am not saying to throw away your old hardware as a rule but there is a point where some of this old stuff just isn’t even worth running.
The reason more performance/watt is in demand because a datacenter can't suddenly draw twice as much power.
I have two LARGE Xeon systems of this era that I used to use when I was heavily involved with Kubernetes and needed to build out a home lab. One is 2x Xeon w/ 256 GB of ram, and one is 1x Xeon w/ 512GB of ram. Both are slow as dogs, and both of them take up at least 150+ watts with only one power supply. My 12th gen Intel Nuc is so, so much faster and efficient. I'm recycling the Xeon systems.
Xeon is a group of products with really varying specs. There is no indication of which XEONs. Also new consumer CPUs often have really small internal caches.
You mention lower footprint but then make a cost comparison against Claude subscription pricing.
Claude subscription pricing is a broken way to consider footprint.
Would you consider improving the website's layout? Right now I find it below average quality and very distracting. Whether you are an engineer or not is not really important; great engineers can write horrible text or use a layout that is not ideal, for instance.
We’re not there yet, but the obvious endgame of the present bubble insanity is open models running on local hardware and devices are “good enough” for most use cases. That will completely implode what’s going on at the moment in tech.
Happened to me. CoPilot changing prices prompted me to cancel my CoPilot subscription and install a local coding model running entirely in VRAM. Will call Claude APIs when I get really stuck, but I should be able to handle 80% of my needs with a dumber local model.
For a long time, too. Programming languages rarely change much, techniques rarely change, so I should be able to use said model for I hope at least five years; and if at any time they optimize local models to cram even more intelligence into the same amount of VRAM, I can upgrade to that.
I like this path.
> Will call Claude APIs when I get really stuck, but I should be able to handle 80% of my needs with a dumber local model.
I experiment with all of the local models I can fit into 32GB of VRAM and I have subscriptions to multiple SOTA providers.
The difference between them is very large, unfortunately. The local models can handle small tasks and refactoring mostly okay, but doing anything challenging with them becomes a waste of time. Unfortunately the waste isn’t immediately obvious because they will come back with something that looks like it works, but then on closer examination I need to throw it out and reset them in a usable direction.
This path is aligned with Apple and Google’s efforts. Both have things to sell you other than tokens, unlike the AI startups.
This. OpenAI and Anthropic are ultimately compute infrastructure plays and not really AI. Everyone will have models, they'll have the ability to run them. This is why the GPU shortage is in their favor.
And like Google and Meta, these companies are going to morph into advertising giants. Advertising is an economic black hole and it eats everything that comes close.
Embedding ads in LLM responses is something researchers are having a lot of trouble figuring out right now.
I have seen the results of some early attempts. It fails in such hilarious ways that all these companies are scared of productizing it. But once someone does it, the taboo is broken and everyone else will follow suit immediately.
It's already being done: https://openai.com/index/testing-ads-in-chatgpt/
How does that view align with Anthropic leasing data centers from others?
I don’t know OpenAI’s infra, but to the extent they are buying GPUs and building data centers with their own money, that sounds like a bad move.
Satya has mismanaged the AI transition in many ways, but one thing he got right is that models are commodities, and the value is in applications that apply them to create user benefit. I agree that any company trying to build a moat with a model is not long for this world.
Then they go bankrupt.
Do you think there will still be an incentive to release weights in that scenario? Everyone will have models only if there continue to be companies releasing weights.
Companies won't but I suspect this is a role that something else open source-y will fill that niche. Maybe orgs like wikimedia or internet archive, maybe some hackers just making things, maybe nation states that want to disrupt other players. Also model training will get better and better both on the algo and the hardware side. You can easily see a world where you might be able to train a good enough model on a home lab in a few days.
But you will need training data. Like a whole Internet search engine or massive data scraping. That‘s a thing that will not change with better algorithms, hardware or cheaper energy.
Data is the only moat but they'll be starting in the same place the current set of players statyed out just a few years ago. I suspect that the delta between what is publicly available (if not legally publicly available! see scihub) and what open ai and anthropic have is relatively small.
Maybe. But if we can all run our own model locally in 2 years on commodity hardware OpenAI and Anthropic will start to look like WeWork during the pandemic
I agree with you that they are headed in that direction! The GPU shortage is (I think) similar to the pandemic era hiring binge. It's less about the extra compute and more about denying the GPUs to potential competitors. They're racing against time to find something that gives them real moat (gen ai I guess?) and they are trading money for time.
This is also why the money being poured into datacenters isn't going to result in as much development as you think. It's about leveraging other people's money to lockdown more future hardware. This is going to end exactly like fiber build out in the 2000s. Eventually that fiber got used but the folks who originally paid for it got hosed.
And free model supply will stop…
I wonder if Google will put out a free model with the ads already baked in.
If you mean releasing model weights: They won't, because they know the "shill something" vector will get abliterated immediately. And they can't use trade secrets or copyright to stop it, either, because they released the model themselves and you don't need to redistribute weights, just an adblocker LoRA.
You just described the absolute nightmare scenario for the newly minted trillion-dollar companies whose only hope is for enterprises and SMB to move all their business processes to the cloud, with employees competing at token maxxing.
I wouldn't say "completely implode", too much money was poured int it, but it's clear we're heading in that direction. You get a model that is "good enough", plus privacy, plus savings in the long term.
Paradoxically, the better results we get from general harness of coding agents, the less moat Claude and co. get. It's unbelievably how fast some open models outpaced frontier models of just a few months ago.
I keep intending to find time to try them. What are you seeing the best results with?
If you are willing to spend about 2000 on GPUs, we are almost there.
In my opinion, the bottleneck is the package management layer and not the model capabilities and performance.
I have been an avid Linux user for decades, and if I find it confusing and painful, something is missing.
this is sorta like saying that being able to run your blog on your laptop will completely implode the cloud business
This is actually what happens.
I run my word processing software on my apple 2 (a total joke of a computer) instead of running it on the WANG.
I run my book keeping software on visicalc instead of the IBM.
I run my simulation software on my IBM PC (I even paid for the 8087!) instead of the VAX.
Moore's law has, at least so far, allowed the pioneers with toy computers to grow their toys big enough to solve "big boy" problems after some time has allowed the toy computers to be faster and the pioneers have scaled their crappy home-grown solution to solve their 60% of the problem that was originally solved by some enormous complex system.
Eventually the toy infrastructure gets expensive and solves 90-120% of the "big iron" problem space, but it also grows to cost as much as the big iron solution, but then a new generation of toy software and toy systems emerges to disrupt the "big iron" systems.
See also http://www.catb.org/jargon/html/W/wheel-of-reincarnation.htm...
You're right Moore's law has been holding up, but will hit a hard limit on process node size, so all scaling will be based on multiple cores. OTH, computing per watt spent has been plateauing. If the future bottlenecks are energy and cooling, that will require infrastructure-scale solutions. My bet is this is going to be real AI company moat.
https://www.riq.net.br/pub/computing-scaling/
Under appreciated requirement for this to work in post-cloud times: open source
If a vendor can SaaS a solution, then enterprise is generally happy (they don't want to have to hire folks for maintenance), and that completely locks out any ability to run locally.
Between enterprise's ambivalence and the obvious financial incentive to vendors, you get SaaS-only products.
It's a huge difference. If you had AI sufficiently good running locally on a phone, you could devise workflows for things like basic digital hygiene, technical assistance, and tedious tasks like inbox management, image sorting, device updates, and so on. Privacy and security gets a big boost past some local competence threshold, and we're nearly there.
Make the local AI competent enough to do good image generation and editing, realtime voice and music generation, handle agentic tasks with a framework like Hermes, and you can take your AI places to do tasks in contexts that are inaccessible to or inappropriate for cloud.
Frontier big platform models will be the best, but there's a level of "good enough" for local uses that we're already seeing flourish, and "good enough" for the average joe is almost here.
Phones and laptops are terrible devices for local AI, way too constrained by bad thermals and small batteries. MiniPC's (many of them using mobile hardware) don't have that particular issue, and can easily run on a 24/7 basis.
The same devices that do what would be considered absolutely monstrous amounts of image and other signal processing just a few years ago. And have an always-on radio.
Phones are also a terrible place to run a radio, but there's a huge amount of benefit in figuring out how to do so.
That level of local AI is also more or less what you need for competent autonomous robots, too. If your household robots are orchestrated from your phone, the local security and cloud convenience converge on a single device. No extra servers, etc, reduced cost, all that - local AI is a massive market amplifier.
It's a little different because cloud and blogs didn't actively get in the way of your home compute. To wit, the various cost spikes for hardware.
People -- WANT -- this technology on their home devices and (apparently?) the providers of this tech don't seem to be running a profit so they probably don't want the maintenance tail on their side either.
I think it's a bit different. Inevitable that this becomes a household-run thing? Not likely.
Running an LLM locally is theoretically viable. Running your blog on your laptop is never viable (unless you hook it up like a server). One just requires compute while the other a stable network.
tbh, my home network is pretty close to the stability of my host these days…
But my downtimes are a bit self-inflicted: changing ISPs which I can personally workaround but harder for a blog where one expects uptime.
The primary feature of a blog or any website is that it is available around the clock, that is the primary feature of cloud: around on the clock computer and network that scales on demand.
The primary feature of "AI" is to process information and reason with a natural language interface at speed, the primary feature of AI bigboys is to provide the machinery that runs the "models".
See the difference?
You severely underestimate how little the fraction of the performance and human labor of a frontier AI is in "the model".
Hosting a blog 24x7 on a laptop is trivial, except for hyperscaling to the front page of HN and Reddit.
More like implode proprietary blog hosting platforms and replace them with commodity VMs that can be used for blog hosting, among other things
Wouldn't arcade cabinets vs home video game consoles be a more apt comparison?
You have to consider that the enshittification factor is much higher now than in the cloud-for-free age.
I disagree. We are currently in a weird period where these frontier AI companies are losing tons of money even on the subscription-based AI models. It's just too compute intensive and there's no way most people are going to be buying the kind of hardware required to run $20 worth of inference every day.
Sadly - it's going to be ads. Advertising is going to get in there and enshittify the whole thing because as always, advertising income is too easy and too plentiful for any company to resist.
Right now the models are fairly agnostic, but we are a hair-breadth away from ChatGPT responding with, "the right tool for this job is a circular saw - something like the Milwaulkee M18, which happens to be on sale at Home Depot this weekend."
$20/day x 250 days per year x # devs/agents/etc = $$$. About $5k per dev at that daily use case.
Enough to validate repurposing an existing workstation with enough RAM, or finding a used high VRAM GPU, or in my case buying a Strix Halo system for home lab and local models.
The future is once again not cloud based, for AI tools.
Most people are running a whole lot less than $20's worth of tokens per day on cloud platforms. (Is that assuming a frontier model? 1M output tokens per day?) Local hardware could easily take up that workload, at least the part of it that's non-time-critical.
The advertising future looks like that to me, too. Service proxies like OpenRouter might talk about price optimization, maybe some ad filtering. But I expect proxies will have malicious entries, too, surreptitiously altering agentic prompts.
Ads are usually the workaround where you don’t deliver enough value to get people to subscribe or payments are unavailable for some reason.
It makes sense to show some ads and get some money at low volume (like a faraway reader wanting to read a story in your local newspaper) but taking money from regular users directly will pay much more.
Newspapers are happy to cannibalize 99% of their ad revenue with a paywall if that 1% subscribes because that’s how much more money you make from someone paying $10-$20/month vs ads.
But yeah, if people use it as a buying recommendation engine, that’s where the money is on ads/referrals but a lot of AI use has little/no connection to buying intent touchpoints.
Newspapers had no choice after craigslist and later Google/Facebook took all their classified revenue.
LLMs may or may not be able to cover their costs with it. We'll see - I suspect product placement as recommendations will become a thing as it won't take as much GPU to give a "recommendation" on "the best widget for X". I firmly expect it to become enshittified the same way google and amazon search has.
And that's if LLMs don't become commodified.
For agentic services, how would you be able to tell that you’ve been product-placed?
Hidden advertising is illegal in most jurisdictions, so it has to be indicated to the user for each specific occurrence and hence be trackable anyway.
"AI can make mistakes. Responses include sponsored content or weights."
Now it's compliant with the law.
Curious when NVIDIA monopoly will ends. China will sure release something that can runs on commodity hardware. I wish they will soon.
I find that hard to believe. The AI companies will want to control what's possible and find new things to do that "need" their services. Otherwise it would be like Intel and Microsoft had decided in the year 2000 that computers are "good enough" now and we would have explored what's possible with that hardware ever since.
> Otherwise it would be like Intel and Microsoft had decided in the year 2000 that computers are "good enough" now and we would have explored what's possible with that hardware ever since.
I think you've misunderstood what good enough means in the context - which is a model capable of completing the tasks assigned to it without having the breadth of full generalization. Your analogy breaks down because of this - we did get 'good enough' spec profiles for different hardware. That thing you're wearing on your wrist won't have the same specifications as the box you use to play games.
I think you've misunderstood the analogy. Just ignore it, analogies mostly break down anyways.
> a model capable of completing the tasks assigned to it
The thing is, the "task assigned to it" is changing with improved capabilities. If everyone around you in 2036 is using general AI to do amazing stuff, you will probably have little interest in vibe coding slop like it's 2026.
>The thing is, the "task assigned to it" is changing with improved capabilities.
Only if you give in to fads and FOMO.
The core tasks people need change at a much smaller pace.
Analogies are like metaphors, they’re illustrative rather than literal.
> The AI companies will want to control what's possible and find new things to do that "need" their services.
That's correct. The problem is they have smart people, tons of money, and several years to figure that out, and the best thing they can come up is a coding agent.
That isn’t the best thing they’ve come up with. It’s a marquee product that is fit for public consumption, however.
The ‘best’ things are; - fuzzy pattern matching algorithms for traffic analysis, human and other image target recognition.
- targeting algorithms that identify ‘suspicious’ individuals in large volumes of metadata.
- fraud analysis
- antagonistic image and video generation, both for fooling other fraud analysis, but also for propaganda, screwing with other actors, etc.
- directed high speed content generation (text, pictures, video) to spam the ‘algorithm’ and allow near realtime identification of additional buttons to push for given target audiences.
- massive marketing/ad manipulation.
Those budget line items (and the suppliers) really want to stay off the radar however, as it makes their life harder.
>Otherwise it would be like Intel and Microsoft had decided in the year 2000 that computers are "good enough" now and we would have explored what's possible with that hardware ever since.
That would be the dream... no fucking Electron! No lockdown modules.
Not saying this isn't the case, but my Anthropic subscription costs me less than the electricity would to power such a home inference system.
Gamers Nexus has a good video on this, but if NVIDIA exits the consumer market, and honestly why would they stay when they can charge up to a 100x for the same wafer space for enterprise, AMD would likely do the same. Only Apple really makes consumer hardware suitable for running things locally then, and maybe some weird Qualcomm ARM chip for Windows. It will be hard running things locally if nobody is supplying the hardware.
Nice post and technically impressive work. I agree we need to understand the build pipeline and be able to do things locally. However, depending on your electricity cost, it might not make sense financially. These old servers are not energy efficient at all (I'm guessing that old Xeon server will easily pull 200W on load), and that model is currently at 0.1$/0.3$ per 1M tokens (with 76 tps and 262k context) in Openrouter (also, these servers are LOUD).
EDIT: I stand corrected, 200W is apparently way too high of an estimate. I used to run a bunch of old Xeon servers and they slurped watts like crazy, but I can't remember which ones exactly those were.
2620v4 is not a power slurping beast. Depending on the server board, it might not be either. Servers are often loud, but it depends.
There's a lot of budget hosting built around chips like these, and they're suprisingly power efficient.
It should be closer to 85W on load. And it's incredibly silent on even a low end cooler. I rarely get above 50° Celcius.
OK, then you're in luck. I had a bunch of old 1U rack servers and even in the next room it was too annoying to run them (they had a bunch of 40mm fans which always ran at full speed, because in a server room, no one can hear you scream).
Yeah, 1u is gonna do that. Get something that can accommodate a big tower air cooler such as the Hyper 212 and your airflow will be quieter than the disks.
I don't run it anymore but my old server was a dual xeon (with two of those coolers crammed in) and I rarely heard a peep out of it.
Could it just be really bad cooling? Looking at 9800X3D, it seems like it's running in a similar range wrt TDP unless you really push the 9800X3D. I'm comparing with desktop cpu's because that's what my workload is. cpu governor is set to performance (no schedutil). No audible change in fan speed during heavy compilation or gaming (very silent humming), and i don't have any fans beside cheap intake, cpu and exhaust fans (1 each) + an excessive amount of dust.
These servers had no fan control whatsoever, they always ran full blast. That's not untypical for rack servers, because as written: they are designed for server rooms, and you're supposed to wear ear protection there anyway... Yes, I could've modified them, but I ditched them because running them simply made no sense (especially the high idle power consumption was ridiculous).
85W for the whole system?! The specifications for the CPU mention a TDP of 85W [1].
[1] https://www.intel.com/content/www/us/en/products/sku/92986/i...
But for LLM work the CPU is mostly idle, waiting for new data - so the CPU itself might not pull much power at all.
Only when you remove it from the original server or enable low fan mode (if available). Most 1U/2U cases will happily blow at full speed well over 90db.
You likely need to replace the flow-through server chassis system with an active "normal" cooler to achieve a bit of silence.
85W might be about right. My old server CPU is in the same ballpark and compiling kernels it reached about 90w in power usage. If you want to keep it running: idle is not very low power unless you have one of the "low power" L versions, keep that in mind.
Get a 4U case, many options if you want to combine it with a NAS. Not hard to cool and keep somewhat quiet. If you can store it in a closet or something that helps too.
Well, you can use it for lots of other things as well.
Compared to the cloud you can probably save up to buy a new server every month. And don't underestimate the gains of having something to experiment on and play with.
These servers are loud if you're trying to fit them into a 1U or 2U, which requires high speed fans to generate the necessary static pressure to push air through the case. I run a similar setup in a 4U case with slow 120mm fans and it's fine.
Glad to see other people realizing this. I've been running Gemma 26B-A4B Q4 on a 2012 Xeon with 16GB to 24GB of RAM in a container. It's getting around 8 to 12 tokens per second. Obviously it's not comparable to huge contexts and running it on a GPU and the image decoder in llama.cpp is super slow compared to a GPU but for some small automation tasks and general trivia questions it's decent. The speed is just enough to not have to wait for it to finish so you can read along.
Here's my setup. You may want to figure out what the best optimizations are for your specific CPU like AVX2 because mine didn't have most of them. I did try MTP briefly but I wasn't getting performance improvements. You could play around with the batch sizes for cache or context or go even lower for Q2 and don't overcommit on threads either, but I would suggest either defaults or trying out llama-bench. This isn't by any means the best I assume but it worked decently for me and I sometimes swap out Gemma for Qwen. You could also lower q8_0 to q4_0 for more context but it could hurt quality some say, altough I have noticed it too on some models.
# Building
cmake -B build -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=ON -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DGGML_OPENMP=ON
# Running
export OPENBLAS_NUM_THREADS=4
export OMP_NUM_THREADS=4
OPENBLAS_NUM_THREADS=4 OMP_NUM_THREADS=4 \
llama.cpp/build/bin/llama-server -hf unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q4_K_XL --temp 1.0 --top-p 0.95 --top-k 64 --min-p 0.00 --jinja --host 0.0.0.0 --port 8080 --cache-type-k q8_0 --cache-type-v q8_0 --threads 4 --threads-batch 4 --ctx-size 8192 -n 8192 --batch-size 2048 --ubatch-size 512 --no-mmap --mlock --chat-template-kwargs '{"enable_thinking":false}' --no-mmproj -np 1 -fa 1
I'm setting up a Frankenstein system at the moment. It's a Chinese DDR3 X99 motherboard with a 12 core Xeon v3, 32gb 1866MT/s ram, and a 1080 Ti.
I'm shoehorning it back in the Optiplex that donated the ram, so it's not ready to go at the moment, but when I had it running on top of the motherboard box as a test I ran the (9B?) gemma4:e4b-it-q4_K_M since it can fit entirely in the 11gb vram. It flew, more than 50tk/s. A model that small isn't useful for coding, but there could be uses. I'd love to figure out a Wake-on-Use and use it as my personal ChatGPT. I'm not sure how that would work... Maybe proxy the LLM thru a Pi with a script to Wake-on-LAN the PC? It'll be a fun weekend project someday.
My always-on LLM is the dense Gemma4:31b that's not quite half in GPU on a 12gb 2060. It's really slow, but the quality is great and my use case is an automated queue so I'm not sitting there watching the output. I have another 2060 but unfortunately the PC won't POST with both installed for some reason.
Result is ~12 tokens per second, as reported by OP down in these comments here.
An impressive effort, and better than I would have thought possible on this hardware -- but still pretty far short of what one needs for an satisfactory interactive session.
Especially if you consider those smaller models are really cheap and fast on platforms like openrouter. Often by the factor 100-500 cheaper than SOTA models, and 2-5x in TPS.
Right. You can also perform RSA encryption on pencil and paper with a scientific calculator. It works, but it's not useful throughput for serious work
Yeah took way too long to find that result. Being able to run on slow RAM isn't surprising considering you can run a model off an SSD.
I was about to ask that
What intrigues me the most about AI progress, is not AGI or the model du jour by $AI_UNICORN, but rather what can be run locally. I remember having an amusing, but rather useless model in a beefy gaming PC that I had 6 years ago; and now, something that’s a hundred times better on my M5 laptop.
Should the market react to the memory shortage, the progress of the Apple silicon continue at the same pace, and what we’ll be able to run locally in 6 years will be very exciting. or frightening.
Also I don’t know what this means for the valuation of the AI companies. I remember asking about this very idea to one of their employees at an event and instead of answering he bailed out to grab a cocktail.
Things you are not supposed to talk about:
- There is no "moat" (lasting, easy-to-defend technological edge) in AI model businesses. There are just short-term advantages.
- An AI business is a capital-intensive business, just like old factories. Data centers are expensive, models are energy-hungry, and the hardware inside must be replaced every 3–4 years.
- Smaller, specialized models eat margins from below. Transcription, voice, or image detection do not need large models.
There is no reason to expect high margins like you can in traditional software business. Benefits of AI go mostly to consumers.
edit: There is potential for economies of scale. Few megacorps can strive for cost advantage when they achieve scale (Microsoft, Google, Amazon and Meta)
All true.
It does seem like the structural characteristics we’ve observed so far suggest there is a kind of flywheel from short-term to long-term advantage due to the capital requirements at various levels.
If you’re Nvidia, making the best GPUs today, the expanding wavefront of demand is consuming them with volume and margins to give you a huge edge in building out the best next generation of GPUs. Similar to how the mobile wave gave TSMC sustained advantage for about a decade now.
I’m guessing this is also what we’re seeing as Anthropic and OpenAI swap spots in the token-vendor market.
I can see the fly wheel in action for Nvidia[1], but in terms of model building - I think the companies that have the advantage here are not Anthropic or OpenAI, but rather companies with substantial revenues from other sources - Google is the obvious player here - reported to be planning on spending 185 billion this year without having a raise a dime from the markets, but there are plenty of other companies - like Meta or Alibaba who can easily fund the longer game from existing revenues.
Everybody talks about this stuff all the time
What you can run locally in consumer hardware is progressing pretty well.
If you get a not-quite-the-best gaming GPU like a 5080, you can run local models that are better than the state of the art from early 2025. Depending on what you want to do, you might have to switch models. The one size fits all huge models are still a data center thing.
Its a convenience thing. You can run a whole lot of stuff locally from wikipedia to social media/email/video servers whatever. Most people with a full time job and 2 kids dont do it cause who has time and energy to patch and maintain the ever growing complexity of this stuff. These systems will keep growing complex. That also means more bugs. Age old tradeoff between freedom and convenience.
You can run mediawiki at home but you won't have wikipedia. You can run a video server but you won't have all the movies that Netfix has. A local model is actually the real thing.
you can have the whole wiki loaded with full search available locally. check out kiwix.
--what this means for the valuation of the AI companies
Probably nothing. Most users have no idea what an LLM is or how it runs. Anecdotally speaking, I see many LLM users default to whatever their day job provides to them. And even slightly more sophisticated users seem ok with paying for their openai or anthropic subscriptions.
Maybe we will see a small but dedicated group of open weight model users who prefer local llm, but everybody else will just consume from the big providers? The scenario might look something like OS choices today - a small, committed group of Linux users vs the vast majority of other users running Windows, MacOS, or Chrome?
This has always been true of software, particularly games. You can get a 5-6 year old game for a fraction of the price, and run it on modest hardware. But the industry wont sit on its hands for 5 years, there will be newer software that requires better hardware.
Technology doesn't always work like that.
A new game is a totally new world with everything created from scratch. A creation. A model, on the other hand, is a reinterpretation machine for hundreds of years of human creations, but not a creation in itself, more like a discovery.
You would think that by now we would have a much better Bitcoin that's taking over the payment networks of the world but what we actually got is a shitload of shitcoin.
Training AI models to drive valuation reminds me of high frequency trading
Have to point out one boring thing though: this will use a lot more electricity than newer things. So it'll work, but it'll run up your electric bill.
The E5-2620 v4 is great. Have been using it for 10 years now. Wanted to upgrade until I saw current prices. I have 64 GB ddr4. Paired it with rx 9060 xt 16 GB and games run as fast as ever. Perhaps the cpu is a slight bottleneck in DOOM The Dark Ages, but i'm at 60 fps, so no problem. Light llm on the gpu is a nobrainer, and it's cool to see that things can be tuned to run ok on the cpu. I bought 2667 v4 a month ago for 30$. I'd expect it to give a decent performance boost but I just haven't had the need for it yet, but pushing into llm like in the article I'd probably upgrade because 2667 can handle slightly faster ram.
Intel sacrificing lifetime for short-term gigahertz is a relatively recent phenomenon.
This is among the "real" differences between workstation/server CPUs and commodity chips for laptops/desktops/handhelds.
Even then, if a commodity chip isn't pushed full tilt at all times, and assuming that the venting and dissipation are adequate, a commodity chip can last a long time.
A quick search on Xeon production yields that it goes through a rather rigorous testing. I wouldn't be surprised that server cpu's in a desktop pc works longer. I can't overclock it either, and that probably helps with its lifespan as well. But yeah, the fact that it actually powers on when i click the button and isn't a limiting factor after 10 years is quite something.
You raise two very good points that I didn't think about: (1) better binning/testing, (2) no overclocking. Keep rockin' that elderly Xeon!
>I can't overclock it either
Except you can overclock v3 :)
Back from my old overclocking days - its heat that kills life. And if you keep that under control (what ages is the heatpaste, replace it ever so often) i very much doubt you'll have any life issues from the cpu itself.
Bearings in fans, caps etc. are also stuff that you need to keep an eye on.
I just replaced a i5-660 thats been powered on since 2010 24/7, heatpaste was fucked so it crashed during heavy loads :)
Not my experience.
Apparently Itanium works quite well for LLMs https://medium.com/@tglozar/running-llama-inference-on-intel...
Which makes sense I suppose.
I wish this were somehow tagged with AI, so I would know that it's not about say, general computing or cost-efficiency (e.g. using an old xeon machine from ebay instead of new, in these cost-conscious times.)
As it is, the title is click-bait for me, as 1) it says I need at least a Xeon somehow and 2) as it doesn't say what I actually need it for.
I want to share something strange. I found a typo or two in the post and this absolutely delighted me, because it implies a human wrote the words. (Or was at least heavily involved in the editing.)
Guess I am a species-ist after all ;)
I hope LLMs don’t get trained with this reply and start adding typos for making it look like it came from a human :)
I felt like I had lost something valuable when I switched to mostly AI based programming, because I used to make so many mistakes that the computer would often do truly magical things I did not even realize were possible.
e.g. one time I tried making a collaborative drawing application but I messed up the logic, and the brush strokes would just get temporarily mirrored between the client and server, so you'd see it getting drawn over and over again in a loop.
The drawing wasn't stored anywhere, it existed only in the network packets between client and server. Accidental GNU.
http://www.gnuterrypratchett.com/
So I started working on a tool that adds random errors back into my programs. To reintroduce the possibility of such happy little accidents.
AIs already make typos, not directly intentionally. Since they are token-based, and tokens are lexemes, they can misconjugate works or make grammatical errors.
llama.cpp includes a benchmarking tool called llama-bench https://github.com/ggml-org/llama.cpp/blob/master/tools/llam...
ik_llama includes llama-sweep-bench https://github.com/ikawrakow/ik_llama.cpp/blob/main/examples...
When comparing hardware, the output of these tools is very helpful to let others put it into context. The post says the output is "reading speed" but knowing the prefill and token generation speeds would be a lot more helpful.
Similar recent posting with optimizations for older Xeon:
High-Performance AI on a Budget: Optimizing llama.cpp for Qwen3.5 Inference on a Dual-GPU HP Z440
https://news.ycombinator.com/item?id=47320244
I've got an old HP Z-620 workstation with dual E5-2697 v2 CPUs (24 cores total, 48 threads @ 2.7GHz) and 128GB of DDR3 RAM. The docs say it supports up to 192GB, but I wasn't able to get it to POST with all the RAM slots full.
It's still a "homelab" beast and does great with development and GIS/Mapping applications. I was not able to figure out how to run AI workloads on it with decent performance, however, so I finally broke down and got a dedicated GPU for it. It's pretty great what can still be done with older hardware.
I self host on old HP Z-840 with 2x3.6 GHz Xeons 24 total cores, and 512 GB RAM. Cost me peanuts used and works like a charm for many years already
I may have missed this in the article, but:
What was the net effect of the optimisations? How much faster did it get?
Did some try to estimates what it would take to bake interference for a capable large language model into silicon so that one can pipeline inputs through it and produce outputs at one token per clock cycle?
I'd expect it to require too much RAM bandwidth to be feasible.
RAM is really slow at silicon speeds. Very little is reachable in one clock cycle, unless the clock cycle is abysmally slow.
No RAM. Instead of having a general purpose multiplier that multiplies an input with a weight stored in RAM, just have a multiplier that hardcodes the weight. In some sense replace each weight with a specialized multiplier and wire them together with accumulators and activation functions in between. And some registers for pipelining. If one goes for four bit quantization, one could have sixteen optimized multipliers, one for each possible weight, and the one just selects and connects them according to the model weights and structure.
Example. If you have a neuron with 16 inputs each 8 bit wide and with a 4 bit weight per input, you will have 16 specialized multipliers each scaling its input by the corresponding weight and then the 16 scaled inputs feed into an adder tree and finally an activation function.
The E5 2620-v4 only supports DDR4.
Probably in an x99 motherboard
The memory controller is integrated into the CPU, so the motherboard chipset is irrelevant. There are some OEM-only v3/v4 parts with dual memory controllers, but the E5-2620 v4 is not one of them.
Old hardware is surprisingly effective. I've been considering a side hustle selling offline AI to local businesses who are privacy-sensitive. Medical, legal, places like that.
At the low end, I'd use old Xeons with gobs of DDR3, install some V100s, run a smaller agent for general chat inquiries, and a frontier model for the deeper stuff, with a router that passes between them depending on the complexity.
The frontier model would perform very slowly, but if it's a deep task the user can submit it in a batch in the evening e.g. "Correlate all of these cases and look for patterns" then receive the output with morning coffee.
Of course, AI helped me work out a plan for this. Haha
How about the iMac Pro? Would that work? I was able to put 128gb in it (not as easy as the regular iMac but possible).
I've been running various models on a Mac Pro 2013 (8 cores, 32 GB RAM) at about 8 to 10 t/s for months. It's not fast, but it's more than enough for many actual tasks, in particular background tasks. An iMac pro will do just as well I suppose.
What are the tasks that do well with 8-10 t/s ?
The sort of task you don't expect to end immediately. If extracting data from a bunch of PDFs takes 1 hour or the whole night, that doesn't make much difference to me. It's not fast enough for auto completion and slightly too slow for chat (but bearable IMO).
Running a local llm at 10 t/s overnight to extract data from a few PDFs will burn more in electricity than paying cents for the hosted kimi models.
You can (sometimes) break even if you have a workstation GPU.
Noting for reference that Gemma4 MTP work is in progress[0] on llama.cpp; similar work for Qwen3.6 landed recently and has been great thus far.
[0]: https://github.com/ggml-org/llama.cpp/pull/23398
Doesn't accepting 100% of the MTP draft tokens mean you should just be using the smaller model? Usually the acceptance rate in Qwen36 at least is around 60-70% and the "wrong" tokens are still filled in entirely by the base model, but when you just accept 100% of the draft tokens it seems kind of self defeating unless I'm wrong.
Also I feel like everyone leaves off prompt processing/prefill speeds in these articles. If you are using a very small prompt and asking for mostly generated tokens, sure but I'd love to know the time-to-response of asking for an analysis of an image or a few hundred lines of code.
As far as I know, speculative decoding still verifies that the proposed tokens are what the "big" model would generate, it just uses the guesses to make that process faster. Setting the probability threshold too low then shouldn't affect correctness, just speed (time will be wasted verifying bad guesses).
But won't setting it to accept 100% of the proposed tokens will skip the verification?
None of those settings set the speculative decoder to accept 100% of drafted token. I assume you are looking at --draft-p-min 0.0, if so, you are misunderstanding what it does.
It depends on the type of MTP. If you're using two models, draft + full, then arguably yes, the larger model isn't providing much benefit if you really are seeing 100% acceptance rates. There are other forms of speculative decoding that work within the larger model by itself though, eg. Qwen has additional speculative decoding attention heads, so there is no secondary drafting model.
Does this mean my 15 year old Phenom is too old? But it has 16 gb of DDR3 RAM!
Admittedly web browsers and it don't get along that well. Literally the only thing that drags though on my Slackware 15 system, and even then usually only when it gets to around 15 or so open tabs.
I tried to run gemma 4 on this CPU and it did not go well
https://www.techpowerup.com/cpu-specs/ryzen-7-4800u.c2281
It is way too slow
Very intriguing. This might be the use for my e5-2430 V2 X2 server that's been lying around. DDR3 is (relatively) cheap now too. Could fit 192GB of RAM in it and play around for much cheaper than a new GPU.
I have an old 192GB DDR4 Dell Precision with dual Intel Xeon Gold 6130 that I've considered spinning up. What's giving me pause is 250W at idle.
Surely that number can go lower with some tweaks
I am sure it can. It will still generate a lot of heat when under load.
Are you telling me I should go for it? :)
I do have a dual DGX Spark cluster running MiniMax M2.7 already so I am all for on-prem. But will be interesting how this old machine will perform!
@cafkafk got a recommendation for a good model that fits into 64GB and leaves a couple GB free for other tasks ?
Honestly, at this point you're probably looking at a smaller model, for the Gemma series I'd go with Gemma 4 E4B with drafters, but that's just a hunch from using it on my laptop (where I do have a RTX 4060 M and 96gb ram).
So you'd change the invocation slightly here, but a lot of things you can potentially reuse.
That said, the Gemma 4 E4B models have so far in my experience been... not great when it comes to long context, but they are very passable for basic tasks, and even seem surprisingly okay at tool calls.
Have you tested Qwen3.6 35B? Putting aside the capability claims for that model (which I support, but are not my point here), that 35B has smaller active parameter count than the gemma 4 26B, potentially making both prefill and decode faster out of the box, and has MTP heads built in the model and well supported (you may need to make sure you download a quant that didn't strip them off, as some do to preserve space). I would be curious to see your numbers there too. And if you do test this, please go for a clean one and not a fine-tuned one.
i tried the Q4_K_M model form unsloth with your Q4_K_M drafter, but the required memory to load everything is 72GB. odd. otoh i could load Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled.IQ4_XS.gguf and it requires just ~18 GB:
~/ik_llama.cpp[main]$ build/bin/llama-cli --model ~/models/Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled.IQ4_XS.gguf --spec-type mtp --draft-max 3 --draft-p-min 0.0 --spec-autotune -cnv --color --jinja --special -smgs -sas -mea 256 --temp 0.7 -t 6 --parallel 6 --cpu-moe --merge-up-gate-experts --flash-attn on --mla-use 3 --mlock --run-time-repack --no-kv-offload . works pretty fast, at about 15 t/s:
llama_print_timings: sample time = 45.28 ms / 404 runs ( 0.11 ms per token, 8921.67 tokens per second) llama_print_timings: prompt eval time = 949.42 ms / 51 tokens ( 18.62 ms per token, 53.72 tokens per second) llama_print_timings: eval time = 24067.08 ms / 400 runs ( 60.17 ms per token, 16.62 tokens per second) llama_print_timings: total time = 242192.55 ms / 451 tokens
so i wonder why the params used by the quantified qwen model use way less memory than the ones of gemma.
What's the best way to apply this to slightly more modern hardware - i.e. 5800XT 32GB DDR4, 9060XT 16GB?
Is this John Siracusa? It sounds like it could be something he’d say…
(He has a fully maxed out “last Intel” Mac Pro and laments the lack of replacement).
I have run llama.cpp on an i7-2600 with a 1050. It's too slow for everyday usage but it's not too slow to make it obvious AI is going to be everywhere and in everything. It's too easy to run.
Would there be any advantage of running this as dual Xeon? The CPUs are $5 and a dual mobo is $50...
More memory bandwidth presumably. Not sure how well the ecosystem handles thread pinning though.
Either they have a E5-2620 V2 from 13 years ago, or they have DDR4, not DDR3. The V3 and V4 only support DDR4.
And this is one of those CPUs which had dual slot motherboards so you can have double the fun (and power bill)
https://pcpartpicker.com/products/motherboard/#s=20028,20029...
I have an ancient DDR3 Xeon that doesn't support any AVX (dual x5690 and 96GB 1333 MHz RAM). You reckon it would even build / run at all?
CPU (2012)
Mainboard GPU Memory: 32GBThis works.
Loading will take some minutes, but at 96 you can squeeze the model in and have some headroom around like ~10 GB, although depending on the Xeon, you may have to downgrade to E4B instead. Should still work thou.
It may work - depending on your ram speeds it might not even be that much slower.
I run Win 11 Enterprise on an el cheapo spare parts Xeon E3-1275 V2 + 32 GiB DDR3-2133 + Gigabyte GA-B75M-D3H rev. 1.2 (TPM support)
Granite or sapphire rapids are very under rated for MoE inference loads. But you need a GPU for the KV cache.
Plus many boards also support CXL for RAM expansion over PCI 5!
Source: building a hybrid inference business for regulated industry workloads.
I think one overlooked advantage of older Xeon systems is their availability. Many people can experiment with local AI deployments at a fraction of the cost of building a brand-new setup.
ive been doing the same thing. i refactored a old newtek stream machine . its my new favorite thing to do! adding old PCs to my "starcraft" fleet xD
This is great work.
I'd love if anyone knows how I might fare with an old Dell R710 with 2 x Xeon 5600 (12 cores total) and 96Gb of DDR3.
I don’t think it would work as well as there is no AVX or AVX2 on those older CPUs unfortunately.
I wonder what the tokens per second actually are. Yes, it does say "reading speed" but that varies for everyone, no?
That is a very fair point! I just ran a not very scientific benchmark with the system under load, and posted the raw logs in a sibling comment above, but the short answer is that it's hitting 11.94 tokens per second for generation - while it's also being a binary cache and CI build server.
Totally just vibes based, I think it goes up to 20+ tps when it's not under load (and that's me trying to be conservative). For context, reading speed at 250 wpm would be around 5 to 6 tokens per second.
Huh, that's actually not bad at all! Sure, it's not at the speed of a GPU, but still, 20 tps is cromulent for a CPU.
What kind of tokens per second did the op get I saw nothing of this written.
11.94 tokens/sec (from another answer above)
This and the previous one are insanely good articles. Thank you!
I'm now staring at a 10 year old 4U with 256 GB of DDR4 and thinking hmmmmm
Makes you wonder if its possible to squeeze more tps out of a strix halo system using the 16 zen5 cores as well as the gpu.
In general you’re mem bandwidth constrained so cpu vs gpu often ends up similar on APUs
There are ways to trade off compute power for memory bandwidth (like MTP and other speculative decoding approaches). The CPU and GPU would need to be able to share the same cache for this to work. In the Strix Halo case the GPU has a private cache on the GPU die I think, which is the snag.
If you get the inference engine to route the heavy matrix math to the GPU and the speculative drafting to the CPU without choking on latency it's probably gonna be very fast.
Would love to see the benchmarks if someone actually pulls something like that off.
As someone doing this for fun on a windows 11 machine (96gb ram, 5090 24gb) I wonder if I need any flags to keep the model in memory and avoid swapping to ssd?
I use LM studio and qwen3.5 35B - but never figured out if it is swapping or not.
Om am unrelated note, does anyone know a model that can help with this use case:
https://news.ycombinator.com/item?id=48301635
The article talks about using --mlock
Well, lets get started. I have 4 of those machines, and they are Two dual processor. They all had 32GB of ram, so now I have two with 64GB, and two with zero. They all hand stock K5000s, now how two have two cards. I stripped the uni processors ram and video cards, and put those into the dual procs. They have 256Gb SSDs, and two 1TB disk drives. One machine has 8Gb of VRam across two cards. Dual processors are 8Cx2 and 32 Threads. They can easily play 16 videos at once. For AI, I have not found a model that I can get above 3 tokens a second. Not a one.
I also run a Qwen 3.6 moe A4B on old hardware. I set it up with
numactl --membind=1
so it is constrained to one of the memory sticks which speeds up token generation a little.
When you use page up and page down key when reading that blog the first line on the screen is obscured by the floating bar or what ever it is. It is not even needed for reading.
The webpage's layout is just horrible. Scrolling is also non-default - and thus rather annoying; I had to stop after two scroll events. Why do people think they need so much fancy effects or non-standard behaviour, if their alleged goal is to get information across to other people?
Might consider going for even older CPUs which don't have the Intel ME ring -3 thing which is full of backdoors
I appreciate the downvotes without any reasoning. It's a fact that newer Intel CPUs have Intel ME which was not in older CPUs and significantly increases attack surface if you are not living in a five eyes state.
In a server, you have to worry about the ME only if you also have an Intel Ethernet interface, which is connected to a potentially hostile network.
If that is not true, the ME cannot be controlled remotely.
The existence of the ME is much more worrisome in laptops, where the ME can be accessed remotely through WiFi. There, to be certain that there is no way for the ME to be accessed remotely you would have to disconnect or cut the internal antennas and use a USB dongle for WiFi.
I agree with the first part. I think this article by FSF about Intel's ME summarizes the issue https://static.fsf.org/nosvn/blogs/Intel_ME_Carikli_article_...
As for the second part, I am not sure about how living in a five eyes state would mitigate it. What do you mean by that?
As five eyes citizen you have at least some rights on paper and you can appeal to your government, but if you are foreigner these guys can go gloves off without any fear of retribution.
Try analyzing Epstein files and posting about it, they'll give you a proper penetration test of all your devices to see what you found out about their ex employee.
Nowadays even EU citizens migrating away from US cloud providers are a "national security issue".
Isn't the whole five eyes argument moot because member states spy on citizens from the other countries and trade intel with each other?
No need for that charade if you are a foreigner, even from NATO ally.
How old are we talking?
IIRC it is pre-2008.
Now we need someone try run Kimi K2.6 on old Xeon and DDR3. After all these platforms do support up to 768GB RAM.
You can run these on a turing machine. At what point is it not worth it? At some point the energy to generate each token matters. We often seen token per second. I think a missing metric is tokens per kilowatt. That is what really matters.
It’ll work but yield a token per minute. With ancient servers the throughput is the limiting aspect not mem size
> The argument for speculative decoding is stronger on CPU than on GPU.
Uh. Uuuh.
No?
___
Also
> While a GPU has a massive pool of ultra-fast High-Bandwidth Memory (HBM), a CPU relies on small, lightning-fast “caches” (L1, L2, L3) built directly onto the processor chip.
What purpose does the quoting of "caches" serve there? Is this AI writing written by that model running on that host?