These advancements are particularly evident in AI-related tasks, as demonstrated by the B100's proficiency in the GPT-3 175B inference performance benchmark. The Blackwell architecture, named in honor of mathematician David Harold Blackwell, signifies NVIDIA's ongoing evolution in the GPU market, with the B100 model leading its new AI/HPC GPU family. The company's roadmap indicates the introduction of several products within this family, including variants like the GB200NVL with NVLINK, the standard GB200, and the B40 tailored for visual computing acceleration.
In terms of memory technology, NVIDIA's present H200 series utilizes HBM3e memory with a throughput of 4.8TB/s. However, with Micron's HBM4 not expected until 2025, NVIDIA may explore alternative solutions for the Blackwell series, such as Samsung's offerings.
The Blackwell GPUs are anticipated to leverage TSMC’s 3nm manufacturing process, which promises enhanced efficiency. Although detailed specifications are limited, the use of the GB2XX nomenclature suggests an allocation for the GeForce GPU lineup.
Looking ahead, NVIDIA has disclosed plans for an X100 chip, targeted for a 2025 release, which will expand the product range to include the X40 for Enterprise use and the GX200, combining CPU and GPU capabilities in a Superchip configuration. Similarly, the GB200 is expected to follow the B100, incorporating the Superchip concept.
The debut of the Blackwell series is predicted to occur during the GTC Keynote in March 2024, following the pattern established by the Hopper series reveal. While this event may not introduce any GeForce products, the GeForce RTX 50 series launch is currently positioned for 2025. This timeline presents a decision point for GPU consumers evaluating the benefits of the forthcoming RTX 50 series against the existing RTX 40 models.
