According to the roadmap, NVIDIA will launch the next-generation Blackwell GPU architecture in 2025, and blossomed into two: GB100 series cores for high-performance computing and artificial intelligence, and GB200 series cores for gaming and creation.
According to the latest news from authoritative exposure expert kopite7kimi, the GB100 big core will be equipped with 10 groups of GPCs, and each group of GPCs is further divided into 8 groups of TPCs. this is not the same as the layout of GA100 8 x 8, GH100 8 x 9.
Theoretically, each TPC group should still have 2 groups of SMs, and each SM group should still have 128 CUDA cores, that’s a total of 20,480 CUDA cores, which is an 11% increase over the current GH100 cores.
At the same time, the bit width of the HBM memory has been expanded to 8192-bit, which is a full 1/3 increase compared to the GH100, meaning that the memory capacity can be up to 128GB, and will inevitably be upgraded to HBM3.
Of course, as usual, B100 cards will shield some units and memory bit width/capacity.
We’re more interested in the gaming card core GB202, which is said to have the same 12 GPC groups as the current AD102, but with 8 TPCs per group instead of 6.
This means that GB202 will have a total of 192 groups of SMs and 24,567 CUDA cores, a full 1/3 increase!
Even more amazingly, kopite7kimi confirms previous speculation that the GB202 core will really support 512-bit bit-width, again an increase of 1/3 from the current level, which means that the video memory capacity will be up to 32GB, and it will be the new generation of GDDR7.
The last time there was a native 512-bit memory bit-width, it was the AMD R200/R300 era Hawaii GPU cores. As usual, the RTX 5090 won’t be full-blooded, nor will there necessarily be an RTX 5090 Ti.
According to previous claims, the GB202 core will also have a doubled 128MB cache and a core frequency of up to ~2.9GHz, which can actually exceed 3GHz. It will also have four different levels of cores underneath it such as GB203, GB205, GB206, and GB207.