If recent reports are to be believed, NVIDIA’s RTX 50 series could be the company’s first mainstream MCM-based GPU. Not only that, but Blackwell architecture may deliver the greatest performance boost in the company’s history. NVIDIA recently confirmed the release of the Blackwell GPU architecture in 2024. Because these are preliminary rumblings, this post should be taken with a grain of salt.
The NVIDIA RTX 50 GPUs have a new SM design as well as a raytracing denoising accelerator.
Before I begin, I’d like to point out that the Blackwell architecture was intended to be the successor to the Hopper architecture, and the use of the terms Blackwell and RTX 50 interchangeably could mean a variety of things. Hopper’s parallel release in the consumer segment (RTX 40) had the codename Ada Lovelace, so it’s possible that Blackwell will as well. Alternatively, it is possible that Blackwell is the consumer codename (though this is less likely) and that the datacenter side of things will be renamed. According to previous reports, Blackwell GPUs will be manufactured on TSMC’s 3nm process.
The rumour comes from RedGamingTech, who has received new information about the architecture. For starters, Blackwell will have an entirely new SM structure. This is not surprising given that the underlying microarchitecture is shifting to an MCM design with Blackwell. In addition, Blackwell will use a hyperspeed bus to connect the various SM and chiplets. A denoising accelerator will also be included in the ray tracing pipeline (modern path tracing setups only trace a portion of the sequence, leaving the rest to a denoiser), which should result in significantly improved RT performance.
There is still no word on specifications, though the source mentions that various Blackwell GPUs are being considered, and that much of the binning will be determined by how AMD’s current RDNA 3 and future RDNA4 offerings perform. Finally, RGT leaves us with the following teaser: “biggest perf leap in NVIDIA’s history”.
At the time of its release, NVIDIA Hopper was the world’s fastest 4nm GPU and the first to use HBM3 memory. With a net total of 18,432 CUDA cores, it outperformed even the NVIDIA RTX 4090 (which has 16,384 CUDA cores). Blackwell will be a significant generational upgrade over Hopper (as has always been the case). In a previous leak, four NVIDIA Blackwell GPUs were confirmed.
David Blackwell, an American statistician and mathematician who made significant contributions to game theory, probability theory, information theory, and statistics, inspired the NVIDIA Blackwell architecture. In addition, he was the first African American inducted into the National Academy of Sciences. The Blackwell architecture continues NVIDIA’s trend of naming major architectures after prominent computer scientists and mathematicians, and little else is known about it at this time. It is very likely that we will eventually be introduced to the parallel nomenclature that will complete NVIDIA’s first MCM duo (as Ada Lovelace is to Hopper).
