Howdy!
(moved this comment from the noob question thread because no replies)
I’m not a total noob when it comes to general compute and AI. I’ve been using online models for some time, but I’ve never tried to run one locally.
I’m thinking about buying a new computer for gaming and for running/testing/developing LLMs (not training, only inference and in context learning) . My understanding is that ROCm is becoming decent (and I also hate Nvidia) , so I’m thinking that a Radeon Rx 7900 XTX might be a good start. If I buy the right motherboard I should be able to put another XTX in there as well, later. If I use watercooling.
So first, what do you think about this? Are the 24 gigs of VRAM worth the extra bucks? Or should I just go for a mid-range GPU like the Arc B580?
I’m also curious experimenting with a no-GPU setup. I.e. CPU + lots of RAM. What kind of models do you think I’ll be able to run, with decent performance, if I have something like a Ryzen 7 9800X3D and 128/256 GB of DDR5? How does it compare to the Radeon RX 7900 XTX? Is it possible to utilize both CPU and GPU when running inference with a single model, or is it either or?
Also… Is it not better if noobs post questions in the main thread? Then questions will probably reach more people. It’s not like there is super much activity…
I think this warrants an extra post. And the beginners thread is a year old and I guess not a lot of people watch comments there.
I use KoboldCpp and like to recommend that to people who are new to the hobby or don’t own a proper gaming rig. It’s relatively easy to install and you can try it now, without any GPU, and see if you like it. I’d say it’s usable on CPU up to about 13B (with quantized models). Of course it’ll be orders of magnitude slower than a GPU.
I’d say every bit of VRAM counts. So you might as well buy as much as you can afford. And you’ll be able to run more intelligent models. Use one of the VRAM calculators to see what fits in 16GB or 24GB. And if you need that model and context size.
Edit: And mixing GPU and CPU makes everything considerably slower. It’s a trade-off for people with less VRAM. But in case you buy something new, you should try to fit everything into the GPU only.
24GB VRAM will easily let you run medium-sized models with good context length, and if you’re a gamer the XTX is a beast for raster performance and has good price/performance.
If you want to get serious about LLMs also keep in mind that most models and tools scale well across multiple GPUs, so you might buy one today (even a lesser one with “only” 16 or 12GB) and add another later. Just make sure your motherboard can fit 2, and you have a CPU, RAM and power supply that can handle it.
Here’s a good example from a guy who glued two much more modest cards together with decent results: https://adamniederer.com/blog/rocm-cross-arch.html
I’m also curious experimenting with a no-GPU setup.
In my limited experimentation on a 12th Gen Intel® Core™ i7-12700H with 64GB of RAM this is probably not worth it for anything beyond simple usage. You can do it but I get something like 4-5 toks/sec with LLama 3B Instruct and that’s one of the faster models.
i have an XTX. it has a TDP of 400 watts. if you install two of them you’ve basically built a medium-effect space heater. you’ll need shitloads of cooling and a pretty beefy power supply.
performance-wise it’s pretty good. over 100 tokens a second with llama3 and it runs SDXL-Turbo about as fast as i can type.
word of warning, if you run Linux you need to manually set the fan curves. i had to RMA my first XTX because it didn’t spin the fans up and cooked itself. the VRAM reached 115°C and started failing.
That’s crazy - does it not have any thermal protection? I’ve had CPUs overheat and they tend to throttle/shutdown before I’ve had anything damaged.
no, it goes full speed until it dies if the fans don’t work. it is limited to 300W from factory but that’s about it. it pulls insane amounts of power.
from what I understand the 4090 is worse but that’s also much larger so it probably handles the heat better.
