NVIDIA HGX Platform page (HGX specifications table)

VERBATIM 2026-10-04, curl -L -A 'Mozilla/5.0 (Windows NT 10.0; Win64; x64)'. Retrieved during verification of nvidia-b200 / nvidia-b300. RESULT: the live HGX Specifications table now lists ONLY '1. NVIDIA Rubin NVL8' and '2. NVIDIA Blackwell' - the Blackwell column is the 8-GPU HGX B200 system, and its per-GPU figures are given as system totals ('NVFP4 Inference 400 PFLOPS' and 'FP32 1,040 TFLOPS', 'FP64 265 TFLOPS', 'NVLink ... 28.8 TB/s'), not as a per-SXM table. The per-SXM B200/B300 columns this repo relies on are no longer on this page; the datasheets (vb-nvidia-blackwell-datasheet-firstparty, vb-nvidia-blackwell-ultra-datasheet-firstparty) and the Blackwell Architecture Technical Brief (vb-nvidia-blackwell-architecture-brief) remain the authorities for per-accelerator B200/B300 figures. Recorded so a later audit does not treat this page as the primary source for B200 per-GPU numbers. Also carries the Rubin-vs-B200 comparison language ('HGX Rubin NVL8 with Sparse NVFP4, HGX B200 with Dense NVFP4'), which is NVIDIA explicitly labelling a NVFP4 comparison number as SPARSE on one side and DENSE on the other - independent confirmation that the dense/sparse distinction is material for NVFP4 and that vendors quote whichever side suits them.