hard20 pts · first answer countsgpu-computinginterconnectdebugging

One node's link runs at half speed

In a 64-node job, NCCL bus bandwidth tests show 45 GB/s on most node pairs but 22 GB/s on any pair involving node 17. What is the effect on training?

A.Only node 17's gradients arrive laterB.Ring collectives run near the slowest link's speedC.NCCL routes around node 17 automaticallyD.Nothing, since bandwidth is averaged

Discussion 0

Join free to post in the discussion.

One node's link runs at half speed

Discussion 0

Related questions