NVIDIA GeForce RTX 5090’s GB202 Silicon Will Be “Physically Monolithic”, Leaker Claims

Leakers have revealed quite a few details on NVIDIA’s upcoming flagship GPU thus far, but one question that originated from its specs is the sheer number of CUDA cores available. With over 24,000 cores on tap, how big is the silicon going to be? Is NVIDIA using multi-chip module (MCM) like AMD did for RDNA3? Turns out, not quite.

RTX 5090 = 2x RTX 5080 (Literally)

Serial GPU leaker @kopite7kimi has replied to a question on X (Twitter) inquiring on the nature of the Blackwell GB202 silicon, to which the leaker answered the die will be “physically monolithic”. One interesting tidbit of this GPU is that it packs exactly double the core count as the GB203 – which is expected to serve as the RTX 5080. This provided an interesting theory that GB202 may be utilizing interconnect designs that connects two GB203 into one singular GB202 package, according to Wccftech.

The theory is established on the basis of an X (Twitter) post two months ago from Bryan Catanzaro, NVIDIA’s VP of Applied Deep Learning Research. He noted the new Blackwell silicon (possibility GB100 given it was announced earlier this year) uses an interconnect that allows two dies to act as a single GPU, so it is possible that GB202 is using a similar packaging technique to allow NVIDIA to produce bigger silicon without major yield concerns.

Such interconnect designs has been used in the consumer space before, in the form of Apple’s M2 Ultra – which combines two M2 Max chips with the company’s UltraFusion interconnect that allows both chips to function as a single large die to the host system. AMD’s implementation is different, however, where it offloads the non-core components to separate dies, while graphical processing elements such as Compute Units remain on a single GCD (Graphics Chiplet Die).

Pokdepinion: Safe to say RTX 5090 will be ridiculously powerful, and possible even more power-hungry.