Decommissioned NVIDIA V100s and AMD MI50s are fairly cheap, $200 for 16gb and $400-500 for 32gb, for local experimentation. They are also very old. There's an enthusiast community keeping these two cards alive and working with current platforms and models.
Nitpick, but the V100 doesn't support bfloat16. The performance hit is not a big deal if you're fiddling with local models, but the card is on it's way out in terms of hardware features.
The MI50 does support bf16, but not the current edition of AMD ROCm. Vulkan support is good and the MI50 works with most major platforms (llama.cpp, vllm, etc.), but it's not without some pain points like manual recompilation. Fortunately the open source community has already paid most of your way.
The cooling requirements for these cards cannot be understated. A consumer grade GPU may throttle if in a small case without additional fans, but if given the same treatment a datacenter GPU will overheat itself idling. You will need to buy, at least, a bunch of decent 120mm fans to prevent this or invest in some water cooling.
I ultimately went with an AMD MI100 32GB ($950). I'm an AMD fan, current ROCm editions support it, and it was low-fuss to get things working. I'm debating getting a second so I can try out bigger models like qwen3-coder-next.
It's prefill; slow prefill kills agentic workloads dead.
If you have 100,000 tokens at ~150tok/s per the OP, you're looking at:
You have: 100000 / (150/s)
You want: hms
11 min + 6.6666667 sec
Which is quite a wait indeed.I don't think this is a fair characterization of the situation. I use frontier models via API pre-paid tokens every single day, and I can barely rack up $100 per month. The fact that we figured out how to burn double this in 20 minutes is impressive, but I don't think it reflects the reality that many are experiencing right now. There are some exceptionally gluttonous approaches to harnessing LLMs that I think are serving as convenient straw men in these discussions.
Paying for the API will almost always be more economical than self-hosting equivalent infrastructure. I am not against self-hosting, but the article suggests a primarily economic motivation for this effort. If you are consuming fewer than 10^9 tokens per month, I really don't think it's worth your time to try and compete with the hyperscalars. Most of the money is to be found in the integration of this technology with existing businesses.
In case anyone is interested, I’m using PCIE passthrough on a FreeBSD host to a Linux guest with an older Pascal card. It’s worked great and I’ve been thinking about putting a nicer card in there. The SXM route seems great, but I’ve been burned (almost literally because of the heat) by DC components before.
i've ran some multi vendor frankenstein setups before and sometimes it even works, so i'm curious to hear your experience with it.
- In 2017, the v100 was a ~$10,000 GPU. I believe there was a PCI-e version but this is probably so cheap because SXM2 is going to be harder to use;
- A 5090 has 1800GB/s of internal memory bandwidth (compared to 900GB/s in the 9 year old GPU). Of course a 5090 is substantially more expensive;
- A 5090 has ~21k CUDA cores vs ~5k;
- The current $10k NVidia GPU is the RTX 6000 Pro w/ 96GB of VRAM. It has slightly more CUDA cores but it otherwise pretty much just a 5090. This is unsurprising. NVidia uses VRAM for market segmentation.
Consider this: in 5-10 years, the trillions spent on AI data centers will likewise be sold for scrap most likely. That's how short the runway is for OpenAI and Anthropic to recover that investment.
Anyway, I'm kind of impressed the author managed to get this all to work. I don't think it even would've occurred to me that someone had made an SXM2 adapter, particularly because it's not even used anymore. Like props to whoever did that.
sigh
Had to stop there. Annoying. I can't stand AI use for writing. It makes any otherwise great article feel so disingenuous.
Because humans write exactly like this /s
Isn't a rasbpi with 16gb of RAM $300 now?