Nvidia Blackwell GB100 and GB202 GPU specs revealed by tipster
Even though Nvidia's next generation of graphics cards aren't due to hit shelves until early 2025, information about some high-end SKUs has started trickling in. Kopite7kimi, a well-known Nvidia tipster, has now revealed more information about two next-gen flagship GPUs. Nvidia plans to switch up its GPU naming scheme this time.
Instead of sticking with GB1xx for all its SKUs, it will now be GB1xx for top-shelf data centre GPUs and GB2xx for gaming (and other) variants. The GB100 will have 10 TPC (Texture Processing Clusters) in every one of its 8 SMs (Streaming Multiprocessors) for a total of 160 SMs (assuming 2 SMs per TPC). Other GB102 specs include a 512-bit memory bus for its memory. Previous rumours said it will use an MCM design like AMD did with RDNA 3.
Moving on to the GB102, it will feature 12 SMs with 8 TPCs each for a total of 192 SMs. This is a ~33% increase in the SMs over the AD102 (144). It is too early to estimate the number of CUDA cores as that figure could change from Ada Lovelace (128 CUDA core/SM). The leaker doesn't specify the type of memory it will use. GDDR7 is a possibility but Nvidia will likely reserve it for its datacentre SKUs.
That said, the GeForce RTX 5090 (tentative) with a GB202 GPU will probably ship with fewer CUDA cores on a nerfed bus (384-bit, if past cards are to be used as a reference). Nevertheless, early rumours said it could be as much as 2.6x faster than its Ada Lovelace counterparts. This, however, will also depend on which node Nvidia chooses for Blackwell.
A custom version of TSMC's N3 seems likely if there is enough capacity left over. Otherwise, Nvidia may be force to stick with TSMC 4N for GeForce cards and an N3 variant for its data centre products. It can get away with it, too, because there's no high-end RDNA 4 equivalent as competition.
Source(s)
