NVIDIA introduced the Blackwell B200 graphic processor and the GB200 superchip, which can radically change the industry in the field of AI.
New chip Blackwell B200 from 208 billion transistors promises performance up to 20 Petflops on fp4. The combination of two such graphic processors (GPU) with one GRACE processor in the GB200 superchpe is even more impressive, capable of 30 times to increase the performance for the tasks of withdrawing large language models (LLM), while significantly reducing expenses and energy consumption to 25 times compared to previous solutions Hopper H100.
Comparison of the appearance of the Blackwell B200 (left) and the previous chip H100 (right). The Blackwell B200 graphics processor, feeding the B100, B200 and GB200 accelerators, is equipped with a pair of computing crystals with a limited mesh that exchange data with each other through the NVLINK-HBI connection at a speed of 10 TB/s
One of the key innovations was the introduction of a second generation of a transformer engine that doubles computing power, throughput and size of the model due to the use of 4 bits per neuron instead of 8. Significant improvements have been achieved thanks to the new version of the NVLINK switch, allowing 576 GPU to exchange data with speed 1.8 terabyte/s.
Particular attention is paid to scalability: Nvidia announced the solution GB200 NVL72, Integrating 36 processors and 72 GPU in one liquid-cooled rack, providing overall performance in 720 petaflops for teaching AI or 1.4 exaflops for the output, with support for models of 27 trillion. parameters.
Grace-Blackwell (GB200) combines a 72-core ARM processor with a pair of graphic processors with a capacity of 1200 W
NVIDIA focuses on the attractiveness of its solutions for large companies, mentioning that Amazon, Google, Microsoft and Oracle plan to offer racks GB200 NVL72 as part of the services of cloud services.
NVIDIA states that its systems can scale up to tens of thousands of GB200 superchites, interconnected by a network of 800 Gbit/using the new Quantum-X800 Infiniband (up to 144 connections) or Spectrum-X800 Ethernet (up to 64 connections). The total capacity can be 11.5 exaflops FP4.
racks GB200 NVL72
NVIDIA announcement does not affect the novelty in the field of game graphic processors, but emphasizes the company’s emphasis on calculations and AI, while foreshadowing the possible emergence of a new line of RTX of the 50th series based on the Blackwell architecture. NVIDIA also said that the supply of GB200, together with the B100 and B200, will begin in the second half of 2024, but it is still unclear to what volume.