Back to AIBriefs
LaunchDevelopersAI Models

mKernel library slashes GPU communication overhead in AI workloads

mKernel library slashes GPU communication overhead in AI workloadsMarkTechPost

mKernel, a multi-GPU fused kernel library, reduces communication overhead that can consume 43.6% of the forward pass in MoE models. It aims to accelerate distributed training by fusing compute and communication across nodes.

·
12 days ago
mKernel library slashes GPU communication overhead in AI workloads — AIBriefs