GB300 NVL72 wideEP optimization is claimed to deliver 7.2x better performance per TCO than every other AI chip
SemiAnalysis says the copper backplane expands the scale-up domain to 72, allowing each GPU to hold only some experts.
TLDR
SemiAnalysis claims a wideEP inference optimization from vLLM, Inferact and NVIDIA gives GB300 NVL72 7.2x better performance per TCO than every other AI chip. It says the copper backplane expands the scale-up domain to 72, versus eight on Hopper. That lets each GPU hold a subset of experts, reducing weight-loading pressure and increasing arithmetic intensity through weight reuse across many tokens.
Combined views
—
3 Sources, first seen ago
— likes— comments— saves— reposts
