Memory
Thirty-two gibibytes of GDDR7 on a 512 bit interface at 1792 GB/s: five and a half times the bandwidth of the GeForce GTX 1080, and half the H100's.
| 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 bandwidth is in the architecture whitepaper and not on the product page, which states the interface width and stops. This is the one figure on the card that divides cleanly into its 2016 ancestor's: same metric, same scope, same method, both stated by the same vendor.