AI · Hardware · 2026-10-26 07:08
NVIDIA's Blackwell B200/B300 GPUs consist of two dies linked via NV-HBI, making cross-die HBM access slower. DeepGEMM's MegaMoE kernel addresses this with die-local memory and scheduling optimizations, effectively reintroducing NUMA considerations inside a single GPU.
Auto-generated summary for reference only. RSS