Graphics
NVIDIA H100 (SXM)
A data-center accelerator rated at 34 TFLOP/s of double precision and 67 of single, with 80 GiB of memory behind 3.35 TB/s and a 700 W power ceiling.
- Manufacturer
- NVIDIA
- Architecture
- Hopper
- Capacity
- 80 GiB
- Rate
- 3.35 TB/s
- Record status
- approved
Figures
| Metric | Value | Evidence | Basis | Method | Source |
|---|---|---|---|---|---|
| Peak FP32 rate gpu | 67 TFLOP/s | Confirmed | Vendor-rated | Theoretical peak Stated by the vendor Single-precision rate on the general floating-point units. Equal to this chip's double-precision tensor-core rate, which is a coincidence of two different units and not a statement that precision costs nothing here: on the general units, single precision is twice as fast as double. | |
| Peak FP64 rate gpu | 34 TFLOP/s | Confirmed | Vendor-rated | Theoretical peak Stated by the vendor Double-precision rate on the general floating-point units, as distinct from the tensor cores. Approved rather than provisional because the source names the precision, the first floating-point figure in this catalog of which that is true. It remains a peak: no workload sustains it, and the same chip's tensor cores double it for the same precision, which is recorded separately because a different unit produced it. | |
| Peak FP64 rate gpu | 67 TFLOP/s | Confirmed | Vendor-rated | Tensor core peak Stated by the vendor Double-precision rate through the fourth-generation tensor cores, which perform matrix operations rather than general arithmetic. Twice the general-purpose double-precision rate, and available only to matrix operations shaped the way the tensor cores require. It is in a comparability group of its own for that reason: quoting it beside another machine's general floating-point peak would compare a specialized unit with a general one. | |
| Peak FP16 rate gpu | 1979 TFLOP/s | Confirmed | Vendor-rated | Tensor core peak with sparsity Stated by the vendor Half-precision rate through the tensor cores, with structural sparsity — the source's asterisk footnote reads "With sparsity". Conditional on the data, not just the hardware: the figure assumes a sparsity pattern the workload must actually have, and halves without it. The method records that condition, so this rate cannot be compared with any dense figure, which is why it is recorded separately rather than rounded into a headline number. | |
| Geekbench Compute score, opencl gpu | 341 973 | Confirmed | Measured | Benchmark chart average Published independently geekbench-compute 7 , opencl A dated capture of the benchmark publisher’s rolling average for this graphics processor, read from the 25 July 2026 OpenCL chart. Version 7 calibrates a compute result against a baseline of 100,000, so it may be compared only with another Geekbench 7 OpenCL figure from this curation round, never with a Metal figure, an earlier version or a processor score. The chart’s device string names the memory rather than the form factor; it is read as the SXM module because the PCIe and NVL cards hold rows of their own. | |
| Thermal design power gpu | 700 W | Confirmed | Vendor-rated | Rated power Stated by the vendor "Max Thermal Design Power (TDP): Up to 700W (configurable)" — an upper bound for the module, and one operators may set lower. A cooling-design target for one component, not a machine's consumption, and explicitly configurable downward, so "700 W" is a ceiling that a given deployment may never approach. The metric forbids ratios, which also keeps it away from the whole-system ratings recorded for the PlayStation 2 and the Mac mini: those describe complete computers, this describes a card. |
Fitted to
Editorial notes
It is the first record in this catalog whose floating-point figures state the precision they were obtained in. Every earlier one — the Dreamcast's, the PlayStations', the M1's — gives a rate and leaves the reader to guess, which is why those are provisional and these are not.