GPU Virtualization Scheduling in the Era of Large Models: A Deep Game Between Hardware Genes and Software Strategies
1. Hardware Isolation of Nvidia MIG: From Architectural Design to Engineering Compromise MIG implements 7 independent instances in A100/H100, each exclusively occupying32-512 SM units, 1.8-24GB of VRAM, and independent L2 cache slices. Hardware-level isolation is achieved throughinstance ID tags: the instruction queue, register file, and memory controller access of each SM carry the instance ID, … Read more