NVIDIA's GeForce RTX 5090 flagship GPU is going to feature a massive & monolithic GB202 "Blackwell" die, reports Kopite7kimi.
Based on what we know so far, the NVIDIA GB202 "Blackwell" GPU is going to power the flagship GeForce RTX 5090 graphics card. Preliminary specifications have revealed up to 192 SMs that will account for up to 24,567 CUDA cores if the chip retains the 128 cores per SM design from AD102 "Ada" chips. Now based on a new tweet from insider, Kopite7kimi, it is revealed that the GPU will be using a monolithic design.
While NVIDIA has shifted to a chiplet design for its HPC/AI chips such as the B100 and B200, it will seem like the company still wants to retain the monolithic packages for its consumer-oriented GPU dies. It is said that the GB202 "Blackwell" GPU will be physically monolithic in design and from the earlier reports, we know that it is expected to feature twice the number of SMs and cores as GB203 which is a more cut-down die for the likes of the GeForce RTX 5080. That will form a huge disparity in the performance between the two cards but the RTX 5090 is shaping up to be a complete beast.
We will get to more on the NVIDIA GeForce RTX 5090 in a bit but before that, let's talk a bit more about the GB202 "Blackwell" GPU itself. So it looks like we are looking at a monolithic design but we also can't rule out the possibility of a chiplet-esque design under the hood.
NVIDIA may essentially be packing two GB203 dies on a monolithic package without making it seem like a chiplet design. It will allow for better inter-die communication rather than off-die communication bottlenecks associated with proper chiplet implementations. While NVIDIA has solutions to overcome the bottlenecks such as NVLINK and other interconnects,
Read more on wccftech.com