Search papers, labs, and topics across Lattice.
Affiliation:
1
0
3
Consumer GPUs can now achieve up to 1.66x faster MoE inference by bypassing NCCL's CPU bottlenecks with a novel DMA-driven communication design.
