Memory
Eight gibibytes of GDDR5X on a 256 bit interface, reachable at 320 GB/s.
| Metric | Value | Evidence | Basis | Method | Source |
|---|---|---|---|---|---|
| Memory bandwidth memory | 320 GB/s | Confirmed | Theoretical | Theoretical peak Stated by the vendor Peak transfer rate of the GDDR5X memory on a 256 bit interface. A bus maximum, stated identically by the launch article and the whitepaper. Same metric, same scope, same method as every other theoretical bandwidth here, which is what makes the multiplier against the GeForce RTX 5090 a legitimate one. | |
| Memory bus width memory | 256 bit | Confirmed | Vendor-rated | Design capacity Stated by the vendor Width of the interface between the GP104 and the GDDR5X fitted to the card. Stated in the article’s prose rather than in either specification table, which is why it took a second transcription of a source the catalog already held. | |
| Memory capacity whole-system | 8 GiB | Confirmed | Vendor-rated | Design capacity Stated by the vendor The card's own memory, on a 256 bit interface. No host memory is included. NVIDIA calls it a framebuffer and writes "8 GB", read as 8 GiB in the binary convention this catalog applies to every memory capacity. Recorded at whole-system scope because it is all the memory this card has. As with the H100, it answers a narrower question than a computer's, since work has to be copied here from a host first. |
320 000 000 000 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.
NVIDIA's launch table calls the capacity a framebuffer and the whitepaper does not state it at all; both state the bandwidth. The capacity is read as 8 GiB, the binary convention this catalog applies to every memory part.