Nvidia RTX 5090 GPU may borrow from Apple’s M Ultra chips

GB202 = GB203 x 2. Solve for SLI.

Two Nvidia GPU dies atop bodies of Dragon Ball Z characters in the middle of a fusion dance

SLI may be gone, but it appears GeForce RTX 5090 could see a new GPU fusion of sorts arrive on the scene. Rumours claim GB202, the presumed flagship Blackwell die, is monolithic in design, but leakers now coyly suggest Nvidia isn’t entirely done with multi-GPU setups.

GeForce RTX 5090 should be the best graphics card in Nvidia’s GeForce RTX 50 series, unless the company decides to resurrect its Titan brand of course. As a flagship card, it will pack the largest GPU die of the series, housing all of its class-leading specifications. However, GB202 looks to be an altogether different beast from previous GeForce GPU dies.

An X thread, between kopite7kimi and other X users discussing GeForce RTX 5090 GPU die GB202

In exchanges between various X users, notable Nvidia leaker kopite7kimi had more to say about the GeForce RTX 5090 GPU die. According to them, GB202 is “physically” monolithic but “logically” operates like two GB203 dies. To make matters more confusing, kopite7kimi cryptically shares they “cannot” or “don’t know how to describe” GB202 and suggests we all “wait and see.”

It’s impossible to say with certainty what kopite7kimi’s claims about GB202 mean, but they start to make sense looking at Apple’s M1 and M2 Ultra SoCs, which aren’t altogether dissimilar. From the above, it’s reasonable to conclude that GeForce RTX 5090 features a singular GPU die composed of two smaller interconnected dies. Physically monolithic, logically otherwise.

In an ideal world, this would leave GeForce RTX 5090 with double the CUDA, RT, and Tensor cores of GeForce RTX 5080 but this is entirely dependent on whether Nvidia opts to use the full die in its graphics cards. After all, GeForce RTX 4090 uses a cutdown version of AD102, leaving room for an allegedly cancelled GeForce RTX 40 series Titan.

Previous GeForce RTX 5090 rumours suggest its performance will be 70% faster than GeForce RTX 4090 while also potentially packing an absurd amount of VRAM. Suffice to say, Nvidia is cooking up a seriously awesome flagship if these leaks hold water. Thankfully, we shouldn’t have to wait much longer for Nvidia’s final word on the matter, as the company will helm a keynote during Computex 2024.

We’ll be reporting from the showfloor as events unfold at the Nvidia press conference, so stay tuned for a GeForce RTX 5090 writeup should Jensen pull one from his jacket or oven. In the meantime, take a look at out GeForce RTX 4080 Super review to familiarise yourself with the current state of play.