During SC 2023 celebrations, the green giant took the opportunity to officially present the first details of NVIDIA Blackwell, a next-generation architecture that will enhance its next line high performance graphics accelerators. The B100 GPU will be the big star in this new generation, and it looks like it will have what it takes to make a real breakthrough.

NVIDIA Blackwell will be an architecture designed by artificial intelligence and there is no doubt about it. Graphics accelerators coming to market based on said architecture will be capable work with much more complex tasksand at the same time they will be able to face the burden they represent huge language models, and will also be able to perform highly sophisticated inference and deep learning.

In terms of performance, NVIDIA has stated that Blackwell can offer a significant jump with the GPT-3 model and that it will be possible to achieve up to 178 billion parameters, as we can see in the attached graph. This equates to more than doubling the performance that the NVIDIA Hopper architecture is capable of offering. Jen-Hsun Huang also commented that this new generation will offer a major improvement in memory bandwidthalthough he did not provide specific data.

Rumor has it that this could be NVIDIA’s first GPU architecture to use MCM design, i.e. abandoning the construction of a monolithic core. This would involve using two or more chips to create a super GPU capable of running a whole host of shaders, which of course would translate into more computing power. We can also expect presence The fifth generation of tensor cores.

Before we finish, it is important to clarify one thing, according to the numerous rumors of NVIDIA Blackwell It will be an architecture exclusive to the professional sector as was the NVIDIA Hopper at the time. By that I mean that we won’t see any general consumer graphics cards based on the first and that NVIDIA’s Ada Lovelace will be the successor other different architecture.

Other theories have recently surfaced saying that NVIDIA will use the Blackwell architecture to create mainstream consumer graphics cards and that these will use a total of five different graphics cores, all with monolithic core:

• GB202: GeForce RTX 5090.
• GB203: GeForce RTX 5080.
• GB204: GeForce RTX 5070 Ti and GeForce RTX 5070.
• GB206: GeForce RTX 5060 Ti and GeForce RTX 5060.
• GB207: GeForce RTX 5050.

We still don’t have concrete details at the specification level, but we hope that new information will be released in the second half of next year, as the launch of the GeForce RTX 50 is scheduled for late 2024 or early 2025 at worst.