NVIDIA GB200 NVL72 product page spec table

VERBATIM 2026-10-04, curl -L -A 'Mozilla/5.0 (Windows NT 10.0; Win64; x64)'. Spec table 'GB200 NVL72 Specs1', two columns 'GB200 NVL72' and 'GB200 Grace Blackwell Superchip': 'Configuration | 36 Grace CPU | 72 Blackwell GPUs | 1 Grace CPU | 2 Blackwell GPU' 'NVFP4 Tensor Core2 | 1,440 | 720 PFLOPS | 40 | 20 PFLOPS' 'FP8/FP6 Tensor Core2 | 720 PFLOPS | 20 PFLOPS' 'INT8 Tensor Core2 | 720 POPS | 20 POPS' 'FP16/BF16 Tensor Core2 | 360 PFLOPS | 10 PFLOPS' 'TF32 Tensor Core2 | 180 PFLOPS | 5 PFLOPS' 'FP32 | 5,760 TFLOPS | 160 TFLOPS' 'FP64 / FP64 Tensor Core | 2,880 TFLOPS | 80 TFLOPS' 'GPU Memory | Bandwidth | 13.4 TB HBM3E | 576 TB/s | 372 GB HBM3E | 16 TB/s' 'NVLink Bandwidth | 130 TB/s | 3.6 TB/s' 'CPU Core Count | 2,592 Arm Neoverse V2 cores | 72 Arm Neoverse V2 cores' FOOTNOTES VERBATIM: '1. Specification in sparse | dense. 2. Specification in sparse. Dense is one-half sparse spec shown.' This page is the RACK-SCALE system table and it corroborates the nvidia-gb200-nvl72 record's per-GPU derivations exactly: 1440/72 = 20 NVFP4 sparse, 720/72 = 10 FP8 and INT8 sparse, 360/72 = 5 FP16/BF16 sparse, 180/72 = 2.5 TF32 sparse, 5760/72 = 80 FP32, 2880/72 = 40 FP64 - and half of each tensor figure is the dense value the record stores (10 / 5 / 5 / 2.5 / 1.25 / 40). Memory 13.4 TB / 72 = 186 GB per GPU and 576 TB/s / 72 = 8 TB/s per GPU, both matching the record. NVLink 130 TB/s / 72 = 1.8 TB/s per GPU, matching the record's '~1.8 TB/s per GPU' and the datasheet's per-GPU 1.8 TB/s. So the record is internally consistent with its own source and no rack/per-GPU inconsistency exists. Note the footnote here ALSO lacks the 'All Tensor Core numbers except FP64 with sparsity' clause quoted in the record's notes.