@kopite7kimi, on Twitter, shortly returns to this story of the long run RTX 5090, or at the least its traits. According to him, the cardboard may properly be outfitted with a GDDR7 reminiscence chip.
RTX 5090: GDDR7 reminiscence, in direction of 1.5 TB/s bandwidth?
Of course, the next are solely rumors, and ought to be taken with a grain of salt. However, bearing in mind the newest developments in reminiscence, it’s not unimaginable that the long run 5090 will exploit GDDR7 reminiscence. At least, that’s what the leaker suggests.
Now, by way of VideoCardz, we study that Micron can be providing the sort of chip in its catalog by the top of the yr. If we assume speeds of 32 Bbps, all on a 384-bit reminiscence bus, as is at present the case for the 4090, this offers us a theoretical reminiscence bandwidth of 1.5 TB/s. This represents a 50% enhance on the present 4090! Nevertheless, we are able to anticipate slower reminiscence, and subsequently decrease bandwidth. Let’s simply say it’s a case of all of the planets aligning.
🥹🥹🥹And GDDR7.
— kopite7kimi (@kopite7kimi) November 15, 2023
Otherwise, for the GPU configuration itself, the latter may accommodate as much as 192 SMs, i.e. a complete of 24567 cuda cores. Once once more, that is the chip’s theoretical most configuration. It ought to be famous that, in observe, the RTX 4090 itself doesn’t function a full GPU. In reality, of the 144 SMs that make up the AD102, solely 128 are lively, i.e. 16,384 cuda cores out of a complete of 18,432.
Now, as for its launch, clearly we are able to anticipate it to reach within the second half of 2024. The anticipated launch window could be across the launch of the structure within the skilled sector… But, as soon as once more, that is to be taken with a pinch of salt!
