NVIDIA Blackwell Ultra datasheet (GB300 NVL72 / HGX B300) - the FP4 Sparse|Dense PAIR row and its footnote, read on NVIDIA's own CDN

VERBATIM 2026-10-04, retrieved with curl -L -A 'Mozilla/5.0 (Windows NT 10.0; Win64; x64)' and parsed with pymupdf. Path: https://www.nvidia.com/en-us/data-center/gb300-nvl72/ carries the link 'https://resources.nvidia.com/en-us-blackwell-architecture/blackwell-ultra-datasheet', which is a PDF.js viewer whose page contains two dam-cdn AssetLink PDF URLs. This record is created specifically to settle the FP4 footnote question, which is the single densest/sparsest trap on this sheet, and to give the repo a first-party host for it rather than the reseller host. TECHNICAL SPECIFICATIONS table, per-GPU columns 'GB300 NVL72 | HGX B300', verbatim: 'Blackwell Ultra GPUs | Grace CPUs | 72 | 36 | 8 | 0' 'Total FP4 Tensor Core1 | 1,440 PFLOPS | 1,080 PFLOPS | 144 PFLOPS | 108 PFLOPS' 'Total FP8/FP6 Tensor Core2 | 720 PFLOPS 72 PFLOPS' 'FP4 Tensor Core1 | 20 PFLOPS | 15 PFLOPS 18 PFLOPS | 14 PFLOPS' 'FP8/FP6 Tensor Core2 | 10 PFLOPS 9 PFLOPS' 'INT8 Tensor Core2 | 330 TOPS 307 TOPS' 'FP16/BF16 Tensor Core2 | 5 PFLOPS 4.5 PLFOPS' 'TF32 Tensor Core2 | 2.5 PFLOPS 2.2 PFLOPS' 'FP32 | 80 TFLOPS | 75 TFLOPS' 'FP64/FP64 Tensor Core | 1.3 TFLOPS | 1.2 TFLOPS' 'GPU Memory | Bandwidth | 279 GB HBM3E | 8 TB/s 270 GB HBM3E | 7.7 TB/s' 'Max Thermal Design Power (TDP) | Configurable up to 1,400 W | Configurable up to 1,100 W' 'Interconnect | Fifth-Generation NVLink: 1.8 TB/s PCIe Gen6: 256 GB/s' [sic: '4.5 PLFOPS' is NVIDIA's own typo in the printed sheet, reproduced here unaltered.] THE TWO FOOTNOTES, VERBATIM, AND THIS IS THE WHOLE POINT OF THIS RECORD: footnote 1: 'Specification in Sparse | Dense' footnote 2: 'Specification in sparse. Dense is ½ sparse spec shown.' WHICH CONVENTION APPLIES TO WHICH ROW, ROW BY ROW: (a) FP4 Tensor Core1 carries footnote 1, which is the LABELLED PAIR convention. In the GB300 NVL72 column the printed cell is '20 PFLOPS | 15 PFLOPS' and in the HGX B300 column '18 PFLOPS | 14 PFLOPS'. Reading the pair left to right as sparse-then-dense gives dense FP4 = 15 PFLOPS on GB300 and 14 PFLOPS on HGX B300. It is NOT 'dense is one-half'. Halving this row would be WRONG. (b) EVERY OTHER Tensor Core row on this sheet carries footnote 2, the sparse-only convention: printed sparse, dense = one-half. So dense FP8/FP6 = 5 PFLOPS (GB300) / 4.5 PFLOPS (HGX), dense FP16/BF16 = 2.5 / 2.25 PFLOPS, dense TF32 = 1.25 / 1.1 PFLOPS, dense INT8 = 165 / 153.5 TOPS. (c) FP32 and FP64/FP64 Tensor Core carry NO footnote marker at all. They are dense as printed and must NOT be halved. FP32 80 TFLOPS (GB300) / 75 TFLOPS (HGX); FP64 1.3 / 1.2 TFLOPS. (d) The system-level rows 'Total FP4 Tensor Core1' (1,440 | 1,080 PFLOPS on GB300, 144 | 108 PFLOPS on HGX) also carry footnote 1 and are also pairs. UNITS DISCIPLINE, WHICH IS WHERE THE DENSITY TRAP BITES HARDEST: 20 PFLOPS = 20,000 TFLOPS, and the footnote-1 dense partner 15 PFLOPS = 15,000 TFLOPS. A record that stores 'fp4: 15000, dense: true' for GB300 is CORRECT ONLY IF the number came from the second cell of the pair (15 PFLOPS), not from halving 20. Storing 15000 with dense:true is indistinguishable on its face from a units slip that read '15 PFLOPS' as '1,500 PFLOPS' and multiplied by 10, and the two errors have very different fixes, which is why the row-level reasoning is recorded here. INT8 ROW IS INTERNALLY INCONSISTENT AND IS NOT RESOLVED HERE: the sheet prints 'INT8 Tensor Core2 330 TOPS | 307 TOPS'. If the footnote-2 rule held, dense INT8 would be 165 TOPS, i.e. 33x the dense FP8 figure of 5 PFLOPS on the same table - an implausible 33x step from INT8 to FP8 in one direction while FP4 is only 4x FP8. The most likely reading is a printing error in the sheet (330/307 look like they may belong to a different quantity or a different unit), but that is inference, not a quotation, so the ambiguity is recorded rather than resolved. Neither accelerators/nvidia-b300 nor accelerators/nvidia-gb300 currently carries an INT8 row at all; if one is added it must carry this caveat. COLUMN TRAP: the per-GPU table has two columns. GB300 NVL72 is 279 GB / 8 TB/s / up to 1,400 W. HGX B300 is 270 GB / 7.7 TB/s / up to 1,100 W. On the B200 sheet the equivalent two columns are GB200 NVL72 (186 GB, 1,200 W) and HGX B200 (180 GB, 7.7 TB/s, 1,000 W). The two platforms' per-GPU figures differ and the two tables' footnotes differ, so a figure lifted from one column and a footnote rule lifted from another sheet will produce a record that is wrong in two independent ways.