NVIDIA’s next-generation Blackwell GPU has been the talk of the town for some time now. First, we learned that NVIDIA would be opting for an MCM design for its GB100, then that the chameleon was turning to TSMC and its 3nm process for etching. Now it’s the turn of the GB202 to make news. In layman’s terms, this is the chip that will be used in the future RTX 5090!
GB202: hints at its configuration!
In broad outline, kopite7kimi gives a few hints as to the layout of the GB202. The chip could feature a maximum of twelve GPCs, each with eight TPCs. This configuration is quite similar to the current AD102, which in its full version features 12 GPCs, but fewer TPCs: 6.
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
Inevitably, this increase in TPCs per GPC mechanically leads to an increase in the number of cuda cores. Indeed, each TPC has two SMs, blocks of 128 cuda cores each. If we calculate correctly, we’ll find a total of 24,576 cuda cores in the GB202, a 33% increase on the AD102.
Next, the leaker indicates that the memory part would be based on a 512-bit bus, which seems to be in line with previous rumors.
However, just because the GB202 will have 24576 cuda core doesn’t mean that the future RTX 5090 will share the same configuration… No, it won’t. If we look at the current RTX 4090, it’s far from having a full AD102. At present, the card that comes closest is the RTX 6000 Ada, which has 142 active SMs out of a total of 144 that make up the chip. The same applies to memory, where it’s unlikely that we’ll see a 5090 equipped with 512-bit GDDR7.
Finally, let’s not forget that Blackwell’s launch should take place somewhere in late 2024 or even 2025. Until then, things may change.