Search papers, labs, and topics across Lattice.
1
0
3
2
FPGAs can beat GPUs at dynamically allocating computation for LLM inference, thanks to a new architecture that fuses operations, uses mixed precision, and caches KV values on-chip.
