The highly reliable NVIDIA insider, Kopite7kimi, is prepared to provide information about the upcoming flagship gaming GPU known as Blackwell. Kopite reveals that the GB100 Blackwell data-center GPU is expected to incorporate 10 TPCs (Texture Processing Clusters) in each of its 8 SMs (Streaming Multiprocessors). This arrangement implies a potential increase in the overall core count compared to GH100 (8×9), unless NVIDIA opts to modify the CUDA configuration within these TPCs, which seems unlikely at the moment.
Even more remarkable is the hinting of details concerning the GB202 GPU configuration. It is said to include 12 SMs, each with 8 TPCs, totaling 192 SMs, consistent with earlier leaks. To clarify, while the GB1XX GPU family targets data centers, the GB2XX series is explicitly crafted for gaming and possibly workstation uses, promising to drive NVIDIA’s next-generation gaming experiences. Regrettably, specifics regarding the design of Tensor and RT Cores remain undisclosed at this juncture.
This configuration of 192 SMs translates to 24,576 CUDA cores, unless modifications are made in the future. This marks a significant 1/3 increase compared to the AD102 Ada Lovelace GPU, which presently serves as the flagship in the RTX 40 series. However, it’s important to note that, as of now, NVIDIA has not introduced an RTX 4090 Ti, and they have yet to fully leverage the potential of 144 SMs for gaming, with the closest option being the RTX 6000 ADA model featuring 142 enabled SMs.
Even more importantly, Kopite confirms previous speculations about the GB202 GPU featuring a 512-bit memory bus. It’s important to mention that this particular configuration might not be used in the RTX 5090. If we assume that the RTX 50 series is indeed transitioning to GDDR7 memory technology, this would naturally lead to a significant increase in memory bandwidth.
NVIDIA’s roadmap indicated that the successor to Ada Lovelace is scheduled for a 2025 release, although there are rumors suggesting it might arrive earlier. It’s impossible to determine the exact configuration of the RTX 5090 GPU at this early stage. Additionally, some leakers mention that NVIDIA’s GB202 GPU has yet to complete its development, with various potential variants still under consideration. Nevertheless, based on Kopite’s track record and confidence, it appears that NVIDIA is progressing towards finalizing the specifications for its next-generation flagship gaming GPU.
As I mentioned before, GA100 is 8*8, and GH100 is 8*9. GB100 will have a basic structure like 8*10. GB202 looks like 12*8.
— kopite7kimi (@kopite7kimi) September 28, 2023