Question on the "Batching memory-bound processes on a GPU" section - it says "This enables us to reuse parts of the model that we’ve already loaded into the GPU’s SRAM", but the 10 GB we are loading is into the HBM, right? How did we overcome the HBM <-> SRAM bottleneck?
More generally, how can we find out the size of the SRAM?
Related discussion on serving finetuned LLMs: https://news.ycombinator.com/item?id=38196661
Thanks a lot for the material Varun, neat presentation with exhaustive computations that make it easy to follow. Question on the serving part: vLLM, Deepspeed, TensorRT-LLM... ? Thanks!
Thanks!
vLLM for quick set up, TRT-LLM for best performance. Both available on https://baseten.co/.
That’s really detailed explanation. Can we do something like this for M1 ultra/M2 ultra/M3 max with large RAM ?
Absolutely. Looks like the M1 Ultra has 800GB/s of memory bandwidth and ~20 TFLOPS of compute.
The same calculations from the post should hold, except with these new values.
great post. could you apply this same framework to optimize training as well?
love the deep dive here
Likely trending on home page since this is directly relevant to LLM costs, i.e., questions like "how much would it cost to rebuild ChatGPT from scratch".
highly relevant question today
:)
Good question. Yes, the 10GB available for batching is in the HBM. In a single forward pass, you move the entire model from HBM -> SRAM exactly once. In a batched forward pass, this is still the case, so you end up doing more compute for the same amount of memory movement.
You can calculate the SRAM as follows: an A100 has 108 SMs, and each SM has 192 KB in SRAM (shared memory, aka its L1 cache) [1]. Multiplied out, this is ~20 MB of total SRAM. This happens to match up with the diagram in the Flash Attention paper [2].
[1] https://developer.nvidia.com/blog/cuda-refresher-cuda-progra...
[2] https://arxiv.org/pdf/2205.14135.pdf