Despite the fact that Nvidia GeForce RTX 5000 graphics cards aren’t expected to launch for a few years, at least, credible rumors relaying details on the Blackwell architecture and the specs of its GPU dies are already doing the rounds. Most recently, new claims suggest the flagship of the series will pack 192 Streaming Multiprocessors (SMs), pointing to a massive CUDA core count of 24,576.
For context, the current flagship Nvidia GPU die, AD102, has 18,432 CUDA cores, which would give GB202 a 33 percent advantage. This increase in SMs is also accompanied by a boost in the number of Texture Processor Clusters (TPCs), from 72 to 96, while Graphics Processing Clusters (GPCs) remain at 12.
All this information comes from trusted leaker kopite7kimi, who also reiterates their previous claim that GB202 will use a 512-bit memory bus for increased memory bandwidth. This GPU die will presumably power a potential RTX 5090, but it’s worth noting that even the RTX 4090 doesn’t use the full power of ADA102.
While these specs are impressive, we’re still unfortunately in the dark in regard to Tensor and RT core counts, as well as architectural improvements made to them. Suffice to say, however, that the GB202 GPU die will undoubtedly power some of the best graphics cards we’ll see in Nvidia’s next generation stack.