You haven’t yet digested the overbidding of the RTX 4090? Prepare for graphics cards RTX 5000 even crazier.
No need to panic, we are not talking about the launch of the new NVIDIA generation right now. Ada Lovelace and the GeForce RTX 4000 still have many months ahead of them.
However, this does not prevent rumors around Blackwell – code name for the following architecture in honor of the mathematician David Blackwell – from multiplying. The latest ones come to us from a regular.
A GB202 “Blackwell” GPU
It is indeed kopite7kimi who gives some details – on X.com, formerly Twitter – about the future NVIDIA generation. Details that we take rather seriously, kopite7kimi is not in the habit of talking nonsense.
In this case, our “informant” first mentions the case of Blackwell in its GPU configuration for data centers. There, the GB100 would have 10 TPCs or Texture Processing Clusters within each of the 8 GPCs or Graphics Processing Clusters. Not clear ? Let’s say that this bodes well for a nice increase compared to the previous generation GH100.
What follows will undoubtedly be a little more telling and should in any case concern us more to the extent that kopite7kimi mentions the case of the GB202 GPU which could well be found at the heart of future “RTX 5090” graphics cards.
No less than 24,576 CUDA cores
By combining the 12 GPCs each with 8 GPU TPCs, this time we are talking about 192 stream multiprocessors or stream processors. This is a clear improvement compared to the AD102 GPU of the Ada Lovelace generation (144).
Another way of looking at things, these 192 SMs are equivalent to 24,576 CUDA cores and, here again, the progress is impressive: practically 35% more CUDA cores than what the AD102 offers (18,432 cores). However, it is worth remembering that no RTX 4000 card uses the full potential of the AD102: the RTX 4090 has “only” 16,384.
Kopite7kimi has another important piece of information: this time it does not directly concern the GPU, but rather the memory interface. The bus used by NVIDIA on the GB202 GPU would actually be 512-bit, compared to the 384-bit bus used on the AD102. That said, on the memory side, there are still many questions as NVIDIA seems set to use the new GDDR7 from Samsung.
If we are undoubtedly not at the end of our troubles regarding the rumors around the GeForce RTX 5000 and their spearhead “5090”, everything suggests that NVIDIA is currently finalizing the technical specifications of its future high-end GPU.
Source : VideoCardz
20