I plan to infer 33B models at full precision
At full precision as in FP16, you are not going to be able to fit ot in a 4090. So if that's your goal, between the choices you are giving, there is only 1 choice. But it won't be speedy.
this post was submitted on 24 Nov 2023
1 points (100.0% liked)
LocalLLaMA
1 readers
1 users here now
Community to discuss about Llama, the family of large language models created by Meta AI.
founded 10 months ago
MODERATORS
If you must run at high precision... the best system in that budget is probably a compromise?
Grab a 3090 or 3060 and slap it on the most RAM bandwidth you can get, with a more modest CPU. The GPU will offload prompt processing and enough response layers to help.
If I had the money, I'd go with the cpu.
Also, I'm not sure a 4090 could run 33B modes at full precision. Wouldn't that require like 70GB of vRAM?
at full precision
Full precision is not as useful as you think. Even at 4bit, the losses are not that large.
70B
What is your motivation for such large models? You're sacrificing a lot of speed for the larger model.
Unquantized, 34B models require at least 65GB plus extra depending on context. I cannot see how your comparison of alternatives can work.
You would need 3 x 4090 to run the model.
The Motherboard and RAM alone cost another $2300... The CPU, Ram, and Motherboard looks like $4500 rounded down... How does this make sense versus three 3090s?
I tried a 70B model on my 48 core epyc with 512gb RAM and it was unusable. I think 1.5t/s or so? Even if you double that it's not great. My M1 Ultra runs it comfortably at 6-7t/s and sips power.
Probably a dual 3090 setup would be the most cost effective solution at the moment while the M1/M2 ultra are the most power efficient solution.
What is the frequency of your ram? DDR4 2666 mhz, DDR4 3200 mhz or DDR5 4800mhz? and how are they installed on the motherboard? 4x128gb(quad channel) or 8x64gb(octa channel)?
Rams frequencies are the most important for llm token generation, as these are often the bottleneck. With a 32 or more-core Epyc 7003 cpu in octa channel (DDR4 3200), you can expect 3 to 4 tokens(70b) equivalent to a 200GB/S vram of speed.
For OP, a 48-core cpu (genoa) or more, in 12-channel DDR5 4800 can expect to go up to 6 to 8 tokens(70b) equivalent to 400GB/S vram of speed.
an rtx 4090 is around 1000GB/s but with only 24gb vram, gpus are generally much faster for prompt processing than PC cpu (over 100 times faster), but I don't know about modern server cpu(Genoa), normally they are faster in prompt processing than PC cpu, as they natively support fp16/BF16 operation.
But take it with a pinch of salt, as I don't have these configurations at hand, so you'll have to ask someone who does.
In my case it's an Epyc 7642 with 8x64GB DDR4 2666, so that may be why my generation is significantly slower.
I find anything below 5 tokens per second not really usable, so that's why I stick with my M1 Ultra. It has plenty of really fast RAM and that again explains most likely why it performs so well, if LLMs are that dependend on fast memory.
I also have a 3090 in another machine but that's also just 24gb and I don't want to shell out more money right now for playing with LLMs, if the M1 Ultra is doing good enough :)
If you like m1 ultra, wait until m3 ultra … . M3 max already smokes M1 Max by 3x the speed in inference.
So expect m3 ultra to be in the 20t/s range
For sure! But the M1 ultra still holds up really well. I doubt I will replace it for another 3 years at the very least. Currently CPUs are progressing at an impressive rate across the board. Would I like an M3 ultra? Sure, but do I really need it? Sadly no :) The upgrade to an M5 ultra will be insane though.
4090 bc it will keep its value unlike the CPU
My AMD 7950X3D ( 16 core 32 threads), 64GB DDR5, Single RTX 4090 on 13B Xwin GGUF q8 can run at 45T/S. With exllamav2, 2x 4090 can run 70B q4 at 15T/s. Motherboard is Asus Pro Art AM5. In Local LLama, I think you can run similar speed with RTX 3090s. But in SD, 4090 is 70% better though.