NVIDIA Blackwell Ultra datasheet doc 4169750 (OCT25) - GB300 NVL72 per-GPU and system specs, per-row sparse/dense footnotes

RETRIEVED AND RE-TRANSCRIBED VERBATIM 2026-10-04. Same reseller-hosted copy of NVIDIA doc 4169750 already carried as source nv-blackwell-ultra-datasheet; recorded as a separate source because this pass read the Technical Specifications table row by row specifically to settle the GB300-vs-B300 per-row footnote question, which is the question the new accelerators/nvidia-gb300 record exists to answer. AUDIT FLAG, unchanged from the parent source: the document is NVIDIA-authored but served from a reseller CDN, not an nvidia.com domain; the nvidia.com resource link is recorded as archived_url and is a gated form, not an open copy. VERBATIM, 'Technical Specifications' table, system rows (GB300 NVL72 then HGX B300): 'Blackwell Ultra GPUs 72 | 8', 'Grace CPUs 36 | 0', 'CPU Cores 2,592 Arm Neoverse V2 Cores | -', 'Total FP4 Tensor Core[1] 1,440 PFLOPS | 1,080 PFLOPS', '144 PFLOPS | 108 PFLOPS', 'Total FP8/FP6 Tensor Core[2] 720 PFLOPS | 72 PFLOPS', 'Total Fast Memory 37 TB | 2.1 TB', 'Total Memory Bandwidth 576 TB/s | 62 TB/s', 'Total NVLink Switch Bandwidth 130 TB/s | 14.4 TB/s'. VERBATIM, 'Individual Blackwell Ultra GPU Specifications' rows (GB300 NVL72 column then HGX B300 column): 'FP4 Tensor Core[1] 20 PFLOPS | 15 PFLOPS' and '18 PFLOPS | 14 PFLOPS'; 'FP8/FP6 Tensor Core[2] 10 PFLOPS' and '9 PFLOPS'; 'INT8 Tensor Core[2] 330 TOPS' and '307 TOPS'; 'FP16/BF16 Tensor Core[2] 5 PFLOPS' and '4.5 PLFOPS' [sic - 'PLFOPS' is a typo in NVIDIA's own PDF]; 'TF32 Tensor Core[2] 2.5 PFLOPS' and '2.2 PFLOPS'; 'FP32 80 TFLOPS' and '75 TFLOPS'; 'FP64/FP64 Tensor Core 1.3 TFLOPS' and '1.2 TFLOPS'; 'GPU Memory | Bandwidth 279 GB HBM3E | 8 TB/s' and '270 GB HBM3E | 7.7 TB/s'; 'Multi-Instance GPU (MIG) 7'; 'Decompression Engine Yes'; 'Decoders 7 NVDEC'; 'nvJPEG Max'; 'Max Thermal Design Power (TDP) Configurable up to 1,400 W' and 'Configurable up to 1,100 W'; 'Interconnect Fifth-Generation NVLink: 1.8 TB/s PCIe Gen6: 256 GB/s'. FOOTNOTES VERBATIM - the whole reason this source exists as a separate record: '1. Specification in Sparse | Dense. 2. Specification in sparse. Dense is 1/2 sparse spec shown. 3. Supported formats provide these speed-ups over H100 GPUs: 2x H.264, 1.25x HEVC, 1.25x VP9. AV1'. PER-ROW READING, which is what the new GB300 record depends on and which is NOT uniform across this table: the FP4 row carries FOOTNOTE 1 and is printed as a SPARSE|DENSE PAIR (20 sparse, 15 dense in the GB300 column; 18|14 in the HGX B300 column) - so 15 is dense and must NOT be halved. EVERY OTHER TENSOR ROW (FP8/FP6, INT8, FP16/BF16, TF32) carries FOOTNOTE 2 and is printed SPARSE-ONLY - so dense is exactly one half (10 -> 5, 5 -> 2.5, 2.5 -> 1.25). FP32 and FP64/FP64 Tensor Core carry NO footnote marker at all and so are read as the plain dense figures they are printed as (80, 1.3). TWO FIGURES ON THIS PAGE THAT ARE INTERNALLY INCONSISTENT AND ARE NOT USED FOR ANY FLOPS ROW HERE: (a) the INT8 row '330 TOPS' against an FP8 row of '10 PFLOPS' on the same GPU is off by a factor of ~30, which cannot be a units statement about the same underlying MAC rate - carried as an explicit dispute on the existing nvidia-b300 record and deliberately not propagated; (b) the FP64/FP64 Tensor Core per-GPU 1.3 TFLOPS, which x 72 = 93.6 TFLOPS, does not match the GB300 product page's 100 TFLOPS system figure (source vb-nvidia-gb300-nvl72-spec-table) - a 6% gap left open rather than reconciled. WHAT THIS PAGE DOES NOT PUBLISH: no memory-bus width in bits for the Blackwell Ultra HBM3E interface (only the 8 TB/s aggregate), no per-pin rate, and no rack power figure.