UC Berkeley's UCCL team releases mKernel, a fused CUDA kernel library that merges GPU communication and compute into a s...

UC Berkeley's UCCL team releases mKernel, a fused CUDA kernel library that merges GPU communication and compute into a single persistent kernel. Communication can consume over 40% of AI training time - this approach aims to eliminate that bottleneck. https://www.marktechpost.com/2026/05/29/meet-mkernel-a-multi-gpu-multi-node-fused-kernel-library-for-gpu-driven-communication/ #AIagent #AI #GenAI #AIInfrastructure

Read Original

Related