On the sidelines of NVIDIA’s keynote from Computex 2024, new details have arrived regarding the NVIDIA GeForce RTX 50 “Blackwell” GPUs have been revealed by Kopite7kimi. The well-known insider, in fact, speaking of an alleged lack of interest in Jensen’s presentation, made it known that the NVIDIA Blackwell GPUs will have the same number of GPCs (Graphics Processing Cluster) as the Ada chips. The only exceptions should be the entry-level SKUs with the GB207 chip.
As expected, the NVIDIA GeForce RTX 50 “Blackwell” graphics cards will feature new dies based on the latest architecture and the company will offer different SKUs, each aimed at a different market segment. But let’s get into the specifics of Kopite7kimi’s statements.
NVIDIA Blackwell SKUs
With a series of posts on X, Kopite7kimi provided new information on the design of the new SKUs, in particular the number of GPCs. And according to the insider, NVIDIA will use the same number of GPCs for Blackwell GPUs as the existing Ada lineup.
Lachronic GPC designates a processing unit within a Graphics Processing Unit (GPU) that includes several components essential for graphics processing, including Streaming Multiprocessors (SM), Texture Processing Clusters (TPC), raster units, and other elements functional. In essence, GPCs are building blocks of modern GPUs and represent a crucial part of the GPU’s architecture, contributing to its processing capability and handling of complex graphics operations. A larger number of GPCs in a GPU generally indicates greater processing capacity and superior performance, as it can handle a greater number of rendering operations simultaneously.
But getting back to Blackwell, it was previously reported that NVIDIA’s GB203 GPU would be half the size of the GB202, so essentially, the GB202 would be made up of two GB203 dies. This does not appear to be the case, thus confirming a pure monolithic design for the RTX 5090.
The new data should therefore lead us to the following configurations:
- GB202 (successor to AD102) -> 12 GPC (8×2 SM)
- GB203 (successor to AD103) -> 7 GPC (6×2 SM)
- GB205 (successor to AD104) -> 5 GPC (TBD)
- GB206 (successor to AD106) -> 3 GPC (TBD)
As for the number of SMs, NVIDIA had 128 cores per dual-SM unit in its Ada Lovelace GPUs. The flagship AD102 had a total of 12 GPCs and 6×2 SMs per GPC for a total of 144 SM units. It appears that the GB202 will offer an increased number of SM drives with 8×2 SM per GPC up to 192 SM. We don’t know if NVIDIA will incorporate 128 or more cores per SM for Blackwell, but for now it looks like each chip will have at least the same GPC count as Ada.
Finally, it is mentioned that only the entry-level GB207 will not have the same number of GPCs as its predecessor AD107. Only the Blackwell GB202 and GB203 GPUs for the GeForce RTX 5090 and RTX 5080 are so far scheduled for launch this year, while the GB205, GB206 and GB207 will only be presented to the general public in 2025.
The first high-end cards are scheduled to launch in Q4 2024, with the 5080 expected to arrive before the 5090, so expect more information in the coming months.
#NVIDIA #GeForce #RTX #GPUs #Blackwell #GPC #count #Ada #Chips
