Hardware · MarkTechPost ·

Meet mKernel: A Multi-GPU, Multi-Node Fused Kernel Library for GPU-Driven Communication

Meet mKernel: A Multi-GPU, Multi-Node Fused Kernel Library for GPU-Driven Communication

UC Berkeley's UCCL team releases mKernel, a library that fuses intra-node NVLink, inter-node RDMA, and dense compute into a single persistent CUDA kernel for GPU-driven communication.

Read the full story at MarkTechPost →