NVIDIA L40 GPU datasheet (Ada) - the part the L40S record cited, retained with its true figures so the mismatch is visible

PART 2 AUDIT, 2026-10-05, the counterpart to w5a-part2-nvidia-l40s-product-page-holds-the-l40s-figures. This document is genuinely NVIDIA's, genuinely an L40 datasheet, and genuinely contains a compute table - it is simply the WRONG PART's table. READ WITH LAYOUT PRESERVED (curl + pdftotext -layout), the table reads verbatim: 'GPU Architecture 48GB GDDR6 with ECC / GPU Memory / Memory Bandwidth 864GB/s / NVIDIA Ada Lovelace architecture-based CUDA Cores 18,176 / NVIDIA third-generation RT Cores 142 / NVIDIA fourth-generation Tensor Cores 568 / RT Core performance TFLOPS 209 / 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** / Form Factor 4.4" (H) x 10.5" (L) - dual slot / Max Power Consumption 300W', footnote '**With Sparsity.' THE L40/L40S DELTA, which is why a plausible-looking transcription went wrong. The two parts share the Ada architecture, 18,176 CUDA cores, 142 RT cores, 568 Tensor Cores, 48GB GDDR6 and 864 GB/s, and differ in the compute table and the power: L40 FP32 90.5 / TF32 90.5 / BF16 181.05 / FP16 362 / FP8 362 / INT8 724 dense at 300W, versus L40S FP32 91.6 / TF32 183 / BF16 362.05 / FP16 362.05 / FP8 733 / INT8 733 dense at 350W. An auditor reading only the shape of the table - same page furniture, same 18,176/142/568 header, same 'N | N**' pair notation - and carrying 91.6/183/362/733 across from the L40S page would find every one of those strings absent here. That is the failure this pair of source records documents. Also note the L40 datasheet's own internal inconsistency, which is a further reason not to trust it as a transcription source: its TF32 row reads '90.5 | 181**', identical to the FP32 row, while the L40S page's TF32 row is exactly 2x its FP32 (183 vs 91.6). And its FP16 row (362) is 2x its BF16 row (181.05) while the L40S page makes them equal (362.05 and 362.05). One of those conventions is wrong and NVIDIA does not footnote the discrepancy. Only the L40S page is cited for L40S values.