NVIDIA H100 Tensor Core GPU product specifications

H100 SXM spec table used for the roofline balance point in this slice: BF16 Tensor Core 1,979 TFLOPS, FP8 3,958 TFLOPS, TF32 989 TFLOPS, GPU memory bandwidth 3.35 TB/s. Tensor Core figures are marked '* With sparsity' on the page - dense bf16 is 989 TFLOPS, i.e. the sparse number is 2x dense. H100 NVL differs (1,671 bf16 sparse / 3.9 TB/s). Vendor figures; the balance-point arithmetic uses them as given.