Gaming GB202 “Blackwell” GPU using the same node as data-center B100

Kopite confirms B100 and GB202 share the same process.

NVIDIA unveiled its latest Blackwell GPU architecture designed for data centers and AI acceleration yesterday, marking the first step in their new architecture push. Reports suggest the tech giant will follow up later this year with a fresh lineup of gaming GPUs, all built on the Blackwell foundation. Interestingly, the new B100 GPU won’t be straying too far from its predecessor in terms of process nodes. Instead, it will leverage 4NP, an enhanced iteration of 4N, which is a custom TSMC node initially developed for the Hopper series.

According to Kopite7kimi, NVIDIA will not be making any changes to the gaming GPU series, which will also rely on this node. According to the post, this new node is still said to provide a 30% increase in density of transistors. This means that both Ada and Blackwell will use a variant of 5nm node.

The leaker added that Blackwell GB202 gaming GPU will see a change in L1 cache (not to be mistaken with L2 cache) which should see a significant improvement to AD102 and GA102,both featuring 128 KB size. Such change may increase the throughout of a single Streaming Multiprocessor, says the leaker.

The GB202 GPU is expected to be the flagship processor for the RTX 50 series. Speculations hint at the possibility of it boasting 192 Streaming Multiprocessors (SMs) and a potentially robust 512-bit memory bus. Given that the B100 features a dual-die design, it’s plausible that a similar approach might be taken with the smaller GB202 gaming processor. Further rumors suggest that GB203 specs could be half of what GB202 has to offer, which means approximately 96 SMs and half the memory bus width compared to the GB202 model.

The specs for the remaining Blackwell gaming GPUs remain unclear, however it was stated that all RTX 50 series should support PCIe Gen5, DisplayPort 2.1 and have GDDR7 memory on board.

Source: Kopite7kimi