To get 3-5 tokens a second on 120GB models requires a minimum of 360-600 GB/s throughput (just multiply the numbers~), likely about 30% more due to various inefficiencies, as you usually never reach the maximum theoretical RAM throughput and there are other steps to evaluating the LLM than just the matmuls. So 468-780 GB/s.

This might be what you’re looking for, as a platform base:

https://www.gigabyte.com/Enterprise/Server-Motherboard/MZ73-LM1-rev-10

24 channels of DDR-5 gets you up to 920 GB/s of total memory throughput, so that meets the criterion. About as much as a high-end GPU, actually. The numbers on genoa look surprisingly good (well, maybe not the power consumption; ~1100W for CPU and RAM is a lot more than the ~300W the A100 would use, you could probably power limit it to 150W and still be faster.).

Of course, during prompt processing, you’ll be bottlenecked by the CPU speed. I’d estimate a 32-core genoa CPU does ~ 2 tflops or so of fp64 (based on 9654’s number of 5.4 tflops, it’ll be a bit more than a third due to higher clock speed), so perhaps 4 tflops of fp32 (fp16 I don’t think is native instruction yet in genoa afaik, and fp32 should be 2x of fp64 using AVX). Compare 36 tflops for the 3090; so it’s going to be 1/5th the speed at prompt processing, which is compute limited (two CPUs), or 1/10th if that’s unoptimized for numa. Honestly, that’s not too bad. But, if you want the best of both worlds, add in a 3090, 4090 or 7900XTX and offload the prompt processing with BLAS, and you get decent inference speed for a huge model (basically, roughly equal or better than anything except A100/H100), and also good prompt processing, as the kv cache should fit in the GPU memory.

As far as CPU prices… . the 9334 seems to range from about $700 (used, quality samples) to $2700 (new), and would have the core count. A bit of a step up is the 9354 which has the full cache size. That might be relevant for inference.