NVIDIA’s RTX 50 series could turn out to be the company’s first mainstream MCM-based GPU – if recent reports are to be believed. In addition, the Blackwell architecture could well bring the biggest leap in performance in the company’s history. NVIDIA recently confirmed that the Blackwell GPU architecture will be released in 2024. As these are very preliminary rumours, this post is primarily a rumor so have that grain of salt ready.
NVIDIA RTX 50 GPUs feature a new SM design and a ray tracing denoising accelerator
Before I start, I would like to point out that the Blackwell architecture was planned to be the successor to the Hopper architecture and the leak using Blackwell and RTX 50 interchangeably could mean several things. Hopper’s parallel release in the consumer segment (RTX 40) was codenamed Ada Lovelace, so it’s possible that Blackwell will also get its consumer counterpart. Alternatively, it’s possible that Blackwell is actually the consumer codename (although admittedly that’s less likely) and the data center side gets a new name. Previous reports have indicated that Blackwell GPUs are manufactured in TSMC’s 3nm process.
The rumor comes from RedGamingTech and they got some new information regarding the architecture. First, Blackwell will feature an entirely new SM structure. This is not surprising given that the underlying microarchitecture is moving to an MCM design with Blackwell. In addition, Blackwell will use a hyperspeed bus connecting the various SMs and chiplets. A denoising accelerator will also be part of the ray tracing pipeline (modern path tracing setups don’t really track the full sequence, they do so partially and a denoiser does the rest), which should result in much improved RT performance.
There still doesn’t seem to be a word on the specs, although the source notes that various Blackwell GPUs are being considered and much of the binning will depend on the performance of AMD’s current RDNA 3 and future RDNA4 offerings. Finally, RGT leaves us with the following teaser: “Biggest leap in performance in NVIDIA history”.
NVIDIA Hopper was the world’s fastest 4nm GPU at launch and the world’s first with HBM3 memory. It featured higher specs than even the NVIDIA RTX 4090 (which packs 16,384 CUDA cores) with a total of 18,432 net CUDA cores. Blackwell will offer a significant generational improvement over Hopper (as has always been the case). Four NVIDIA Blackwell GPUs have already been confirmed in a previous leak.
The NVIDIA Blackwell architecture is based on David Blackwell, an American statistician and mathematician who made significant contributions to game theory, probability theory, information theory, and statistics. He was also the first African American to be inducted into the National Academy of Sciences. Blackwell continues NVIDIA’s trend of naming major architectures after prominent computer scientists and mathematicians, and not much more is known about the Blackwell architecture at this time. It’s very likely that at some point we’ll be introduced to the parallel nomenclature (like Ada Lovelace to Hopper) that will complete NVIDIA’s first MCM duo.