Show HN: Hardware-software co-design for MoE models to bypass NCCL bottlenecks (github.com) 3 points by PJHkorea 2mo ago ↗ HN
[–] PJHkorea 2mo ago ↗ Imagine virtual memory, but instead of using your local storage, you’re pooling a whole bunch of network servers into a single address space
1 comment
[ 2.9 ms ] story [ 14.3 ms ] thread