Rumor | Nvidia Blackwell RTX 50 rumored to offer biggest perf leap in history, company planned GH202 Hopper as contingency in case AMD had faster RDNA 3
Nvidia's Ada AD102-based GeForce RTX 4090 has been leading the GPU performance charts since launch, with AMD's RDNA 3 restricting its wrestling arena to the RTX 4080. The rumor mill hardly sleeps, however, and now we already are getting to know some rough information about Nvidia's upcoming GPU architecture codenamed Blackwell.
Blackwell is named after David Harold Blackwell, an American statistician who is credited with important contributions to the fields of game theory and information theory.
The first references for Blackwell popped up in leaked driver code back in March this year alongside references to Ada and Hopper. It is not entirely clear whether Blackwell would be a high-performance compute (HPC) or consumer GPU.
Paul from RedGamingTech on YouTube claims he has information that Nvidia has already begun evaluating Blackwell SKUs internally. These are not taped-out designs but more like proof-of-concept prototypes.
According to Paul's sources, Nvidia is apparently looking to keep its leads intact going in to the next generation — even if it means putting out a halo card that would be made in limited quantities and priced significantly higher than comparable mainstream offerings.
Blackwell is likely to be based on an entirely new multi-chip module (MCM)-based shading multiprocessor (SM) design. Paul notes that one of the evaluation SKUs is an MCM monstrosity with hyperspeed bus links between various chiplets in the MCM package. The current Grace Hopper CPU+GPU package uses a high-speed 900 GB/s interconnect.
Several leaks earlier this year actually pegged GH100 Hopper to be an MCM, but Nvidia opted for a traditional monolithic die instead. Rumor is that Nvidia would have actually brought GH202 Hopper to the consumer segment in an "emergency, break glass scenario" had the RDNA 3-based RX 7900 XTX surpassed the RTX 4090.
RTX 50 Blackwell will also apparently ship with a denoising accelerator as part of the ray tracing pipeline. Nvidia originally used a Spatiotemporal Variance-Guided Filtering (SVGF) denoising solution for real-time ray tracing. This was supplanted by the Nvidia Real-time Denoiser (NRD) that can work with low ray counts to the extent of 0.5 or 1 ray per pixel (rpp) and can work with diffuse, specular, and infinite light source shadows. NRD is slated to offer 50% more performance than SVGF-based denoising.
It looks like Blackwell will integrate denoising directly as part of the RT pipeline, which should theoretically help with higher RT performance, particularly in path tracing.
In the same video, Paul also talks about how RDNA 4 Navi 41 can be 2x as fast as RDNA 3 Navi 31 with much improved performance per Watt.
Nvidia and AMD are even yet to fully launch the complete Ada and Navi 3x stack, respectively, so these very early rumors about their next gen GPUs should be taken with a pound of salt.
Buy the Gigabyte Aorus GeForce RTX 4090 Xtreme Waterforce GPU on Amazon