NVIDIA L40S GPU product page - where the L40S compute figures actually are, as opposed to the L40 datasheet the record cited

PART 2 AUDIT FINDING, 2026-10-05. data/accelerators/nvidia-l40s attributes all six of its compute rows to nv-l40-datasheet ('NVIDIA L40 GPU datasheet', https://images.nvidia.com/content/Solutions/data-center/vgpu-L40-datasheet.pdf). THAT DATASHEET IS FOR A DIFFERENT PART and contains NONE of the six numbers. METHOD 1 - the cited datasheet, fetched as a PDF and read with layout preserved (curl + pdftotext -layout, so the two-column 'dense | sparse' pair is not collapsed). Its compute rows, verbatim: 'FP32 TFLOPS 90.5 | 181**', 'TF32 Tensor Core TFLOPS 90.5 | 181**', 'BFLOAT16 Tensor Core TFLOPS 181.05 | 362.1**', 'FP16 Tensor Core 362 | 724**', 'FP8 Tensor Core 362 | 724**', 'Peak INT8 Tensor TOPS / Peak INT4 Tensor TOPS 724 | 1448**', with the footnote '**With Sparsity.' The stored values are fp32 91.6, tf32 183, bf16 362, fp16 362, fp8 733, int8 733. A grep for '91.6' and '733' over the extracted text returns ZERO matches. The two documents disagree at every row: the L40 prints FP32 90.5 and FP8 362; the L40S prints FP32 91.6 and FP8 733. METHOD 2 - the L40S product page, extracted independently. It carries BOTH a summary table and a full 'NVIDIA L40S GPU Specifications' table, verbatim: 'FP32 | 91.6 teraFLOPS', 'TF32 Tensor Core TFLOPS | 183 I 366*', 'BFLOAT16 Tensor Core TFLOPS | 362.05 I 733*', 'FP16 Tensor Core | 362.05 I 733*', 'FP8 Tensor Core | 733 I 1466*', 'Peak INT8 Tensor TOPS 733 I 1466*', 'RT Core Performance TFLOPS | 212', 'Max Power Consumption | 350W', footnote '*With Sparsity'. A second inline table on the same page reads 'FP32 | 91.6 teraFLOPS | TF32 Tensor Core | 366 teraFLOPS* | FP16 | 733 teraFLOPS* | FP8 | 1466 teraFLOPS*' - that summary table prints only the SPARSE member of each pair, which is a third convention on one page and a trap for anyone transcribing from the wrong table. TWO PRECISION CATCHES THE SUMMARY TABLE HIDES. The full table gives BFLOAT16 and FP16 as 362.05 dense, while the summary table's 'FP16 | 733' is the sparse member of 366.5, not 733/2. So '733' is simultaneously the sparse FP8 value and (by coincidence) 2 x the sparse BF16/FP16 value. The record's fp16 362 and fp8 733 are both correct as DENSE, but only against the full table. And the stored bf16/fp16 362 should be 362.05 against this document; 362 is the L40 GPU's number and appears to have been carried over. THE DENSE FIGURES, restated so the halving is auditable: FP32 91.6 (no sparse variant printed); TF32 183 dense / 366 sparse; BF16 and FP16 362.05 dense / 733 sparse; FP8 733 dense / 1466 sparse; INT8 733 dense / 1466 sparse. The record's dense: true on every row is therefore correct against THIS document - the error was purely the citation, not the values or the dense flags.