AUDIT: the H100 datasheet's 'H100 NVL' column is a 2-CARD aggregate (footnote 1), so the per-card NVL figures live on NVIDIA's H100/H200 product pages - the stored H100-NVL values match the product page, not this datasheet

PART 2 AUDIT FINDING, 2026-10-05. This is the reason a whole class of 'missing' values turned out to be missing, and it is the exact dense/sparse-plus-pair trap this repo has been bitten by. METHOD - PDF fetched with curl and read with 'pdftotext -layout', because plain web_extract collapses the table's cells into one run and destroys the column structure. With layout preserved the table reads verbatim: Technical Specifications H100 SXM H100 PCIe H100 NVL(1) FP64 34 teraFLOPS 26 teraFLOPS 68 teraFLOPS FP64 Tensor Core 67 teraFLOPS 51 teraFLOPS 134 teraFLOPS FP32 67 teraFLOPS 51 teraFLOPS 134 teraFLOPS TF32 Tensor Core 989 teraFLOPS(2) 756 teraFLOPS(2) 1,979 teraFLOPS(2) BFLOAT16 Tensor Core 1,979 teraFLOPS(2) 1,513 teraFLOPS(2) 3,958 teraFLOPS(2) FP16 Tensor Core 1,979 teraFLOPS(2) 1,513 teraFLOPS(2) 3,958 teraFLOPS(2) FP8 Tensor Core 3,958 teraFLOPS(2) 3,026 teraFLOPS(2) 7,916 teraFLOPS(2) INT8 Tensor Core 3,958 TOPS(2) 3,026 TOPS(2) 7,916 TOPS(2) GPU memory 80GB 80GB 188GB GPU memory bandwidth 3.35TB/s 2TB/s 7.8TB/s(3) Max thermal design power (TDP) Up to 700W 300-350W 2x 350-400W FOOTNOTES, verbatim: '1 Preliminary specifications. May be subject to change. Specifications shown for 2x H100 NVL PCIe cards paired with NVLink Bridge' and the sparsity marker '(2)' on every TF32/BF16/FP16/FP8/INT8 row; '3' on the 7.8TB/s. THE TWO TRAPS, AND THEY STACK: (1) '(2)' is the SPARSE marker: dense is one-half of the printed value. So SXM TF32 is really 494.5 dense, not 989. The stored nvidia-h100-sxm tf32 494.5 / bf16 989.5 / fp16 989.5 / fp8 1979 / int8 1979 are correct AS DENSE, correctly halved - and, importantly, the pages never print the '.5' decimals. The stored values are right; the citation merely does not contain them as printed, which is the benign half of this finding. (2) Footnote 1 makes the ENTIRE NVL column a 2-CARD AGGREGATE. The header literally reads 'H100 NVL(1)' and the TDP row says '2x 350-400W'. So this column's per-card figures are 94GB, 3.9TB/s, TF32 989.5 dense, BF16/FP16 1,979 dense, FP8/INT8 3,958 dense, and FP64 34 / FP64-TC 67 / FP32 67 - NOT the 68 / 134 / 134 / 1,979 / 3,958 / 7,916 the table prints. WHICH MEANS THE STORED H100-NVL VALUES ARE RIGHT AND THE DATASHEET IS THE WRONG CITATION FOR THEM. data/accelerators/nvidia-h100-nvl stores fp64 30, fp32 60, tf32 417.5, bf16 835.5, fp16 835.5, fp8 1670.5, int8 1670.5, vram 94GB, 3,938 GB/s. None of those strings occurs anywhere in this PDF. They are NVIDIA's PER-CARD H100 NVL figures as printed on the H100/H200 product pages (TF32 835 sparse -> 417.5 dense, BF16/FP16 1671 sparse -> 835.5 dense, FP8/INT8 3341 sparse -> 1670.5 dense, footnote '* With sparsity'), and they reconcile exactly as half of the datasheet's PAIRED column: 989/2=494.5 is the SXM figure, while 1979/2=989.5 is the paired-NVL per-card BF16 - the arithmetic chain pair -> card -> halve is three steps, and doing only two of them lands on 989.5, which is the SXM dense BF16 and a plausible-looking wrong answer for the NVL card. The record's own notes already document this reconciliation, so this source record exists to attach the PROOF to it: the datasheet is retained as evidence for the pair basis, and the per-card values need the product page cited alongside it.