Accelerator
104.8 TFLOP/s of single precision, 32 GiB of GDDR7 at 1792 GB/s and 575 W. These are the same three quantities recorded for the GeForce GTX 1080 nine years earlier.
A Palit GeForce RTX 5090 GameRock. It is a partner card; the frequencies and memory figures in this record are NVIDIA's for the reference design.
This record can be compared with the GeForce GTX 1080, the oldest consumer card in the catalog. Both are NVIDIA cards, and both state a single-precision peak and memory bandwidth obtained by the same methods. Their multipliers are therefore legitimate: twelve times the floating-point rate, five and a half times the bandwidth, and four times the memory. Board power is a rating rather than a measurement, so the metric forbids a ratio for it.
NVIDIA's page for this card states no floating-point rate or memory bandwidth. It gives "3352 AI TOPS" and "318 TFLOPS" instead. The first counts low-precision tensor operations; the second belongs to the ray-tracing engine. Neither is a single-precision rate. Dividing either by the GeForce GTX 1080's 8,873 GFLOPS would combine different units, arithmetic and precision, so neither is recorded.
The architecture whitepaper supplies the recorded figures. Its appendix writes "Peak FP32 TFLOPS (non-Tensor)" and prints the bandwidth. Like the H100 record, this one relies on a vendor document that names the claimed precision.
| Component | Role | Count | Clock | Capacity |
|---|---|---|---|---|
| NVIDIA GB202 (GeForce RTX 5090) The base clock. The boost clock, 2407 MHz, is a separate figure in a separate group. | GPU | 1 | 2.01 GHz Nominal clock 2407 MHz Peak clock | — |
| NVIDIA GeForce RTX 5090 frame buffer | Video memory | 1 | — | 32 GiB |
| Metric | Value | Evidence | Basis | Method | Source |
|---|---|---|---|---|---|
| Memory bandwidth whole-system | 1792 GB/s | Confirmed | Derived | Theoretical peak Computed here from figures the vendor stated One pool of GDDR7 on the card, and nothing else on it. Adds 1 memory-bandwidth figure obtained by theoretical-peak. Nothing in the catalog is left out of it. A total is only as trustworthy as the decision about what belongs in it, so both the terms and the exclusions are named records. |
| Metric | Value | Evidence | Basis | Method | Source |
|---|---|---|---|---|---|
| Clock frequency gpu | 2.01 GHz | Confirmed | Vendor-rated | Nominal clock Stated by the vendor Reference card, as specified in NVIDIA's product specification table. The base clock, from the product page: the whitepaper's appendix states only the boost clock. It is the figure that belongs beside a fixed frequency, and the reason a reader can put this card next to a Commodore 64 without the comparison quietly changing what it means. | |
| Clock frequency gpu | 2407 MHz | Confirmed | Vendor-rated | Peak clock Stated by the vendor Reference card, the frequency NVIDIA guarantees the card reaches when power and thermal headroom allow. Taken from the whitepaper, which states 2407 MHz, rather than from the product page, which states 2.41 GHz. That is the same figure to three digits, and the more precise one is the one recorded. A ceiling reached when there is headroom, so it shares a comparability group with the GeForce GTX 1080's boost clock and with the PlayStation 5's variable-frequency maximum, and none with a fixed clock. | |
| Peak FP32 rate gpu | 104.8 TFLOP/s | Confirmed | Theoretical | Theoretical peak Stated by the vendor 21 760 cores at the 2407 MHz boost clock, on the general shader units rather than the tensor cores. A theoretical peak, not an achieved rate. The whitepaper writes "Peak FP32 TFLOPS (non-Tensor)", which names both the precision and the units that produce it, so this rate is comparable with the GeForce GTX 1080's and with nothing else on the card. The product page's "3352 AI TOPS" and "318 TFLOPS" are not recorded anywhere in this catalog: the first counts low-precision tensor operations, the second is the ray-tracing engine's, and neither may be divided by a single-precision peak. | |
| Geekbench Compute score, opencl gpu | 335 760 | 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. | |
| Rated power consumption gpu | 575 W | Confirmed | Vendor-rated | Rated power Stated by the vendor "TGP (Total Graphics Power): 575W" for the reference card, alongside a 1000 W required system power that describes a computer rather than this card. NVIDIA's word here is "Total Graphics Power", not TDP: a board power rather than a cooling-design target, which is why it is recorded under a different metric from the GeForce GTX 1080's 180 W. The two are the same kind of claim about two cards and not the same quantity, and neither metric permits a ratio, so the difference costs nothing and hiding it would. | |
| Transistor count gpu | 92 200 000 000 transistors | Confirmed | Vendor-rated | Die specification Stated by the vendor Whole GB202 die. The count is the full die’s. The same whitepaper says the full GB202 carries 192 streaming multiprocessors and that this card has 170, so some of these transistors are not enabled on it. NVIDIA nevertheless prints 92.2 billion in the card’s own column, and the figure is recorded as the vendor states it with the discrepancy named here rather than silently corrected. | |
| Process node gpu | 4 nm | Confirmed | Vendor-rated | Die specification Stated by the vendor Node as named in the whitepaper: "TSMC 4nm 4N NVIDIA Custom Process". "4N" is a foundry process name, not a dimension, and NVIDIA calls it a custom process rather than the stock one, so the four is a label the number 4 happens to sit inside. Recorded because the metric is defined as a marketing node name and forbids ratios outright. |
| Metric | Value | Evidence | Basis | Method | Source |
|---|---|---|---|---|---|
| Memory bandwidth memory | 1792 GB/s | Confirmed | Theoretical | Theoretical peak Stated by the vendor Peak transfer rate of the GDDR7 memory at 28 Gbps on a 512 bit interface. Stated in the architecture whitepaper and nowhere on the page that sells the card, which gives the interface width and stops. Same metric, same scope, same method as the GeForce GTX 1080's, which is what makes a multiplier between the two legitimate. | |
| Memory bus width memory | 512 bit | Confirmed | Vendor-rated | Design capacity Stated by the vendor Width of the interface between the GB202 and the GDDR7 fitted to the card: sixteen 32 bit controllers. One of the few figures NVIDIA prints identically in the whitepaper and on the page that sells the card. | |
| Memory capacity whole-system | 32 GiB | Confirmed | Vendor-rated | Design capacity Stated by the vendor The card's own memory, on a 512 bit interface. No host memory is included. Both documents write "32 GB", read as 32 GiB in the binary convention this catalog applies to every memory capacity. Recorded at whole-system scope because it is all the memory the card has, and it means what a card's memory means rather than what a computer's does. |
1 792 000 000 000 B/s
a + b + … sum v1 Adds 1 memory-bandwidth figure obtained by theoretical-peak. Nothing in the catalog is left out of it. A total is only as trustworthy as the decision about what belongs in it, so both the terms and the exclusions are named records.
The release date is editorial context: neither cited document carries a publication date. The configuration is labeled a reference card for the same reason as the GeForce GTX 1080's. Partner boards carry clocks of their own. The tensor and ray-tracing rates, the fill rates and the required system power are not recorded because the catalog has no metric that would make them comparable with anything. The whitepaper's non-tensor FP16 and INT32 peaks are omitted for the same reason as the H100's BFLOAT16 and FP8 rates.