5 ms·
Rollie calculation, like the online softmax in FA, implies a centralized computing unit that does the compute and stores the intermediate results in its registe
by reliabilityguy 19d ago
Rollie calculation, like the online softmax in FA, implies a centralized computing unit that does the compute and stores the intermediate results in its registers. With PIM you have no centralized compute unit, you have a bunch of memory, and a bunch of MACs all over the place.
How would you do map-reduce across multiple DIMMs w/o extra reads/writes?
PIM implies some sort of distributed compute, which can work for some cases, but I am not sure LLMs are one of them.
- whatshisface 19d agoRe-arrange x+(y+(z+w)) into (x+y)+(z+w) to accomplish a distributed sum. Shared bus, interconnect fabric. This is a supercomputer on a chip.
- ACCount37 19d agoIs there a single reason why we can't just "distribute" the online softmax? Each die-attached PIM accelerator computes online softmax for its own KVs. Then the central unit gathers the softmax intermediates, one intermediate per die, and uses those to compute the final softmax. The PIM win is that we crater the memory traffic between the central accelerator and the memory dies for attention ops. Most of the attention bandwidth never leaves the memory. This isn't "run the entire LLM in PIM", no - this is "offload the parts of LLM that benefit from PIM the most to PIM".