NVIDIA Introduced The Most Powerful Chip In The World – Blackwell B200
Nvidia presented next-generation AI accelerators on GPUs with Blackwell architecture at the GTC 2024 conference. According to the manufacturer, upcoming AI accelerators will make it possible to create even larger neural networks, including working with large language models (LLMs) with trillions of parameters, and at the same time will be up to 25 times more energy-efficient and cost-effective compared to Hopper.
According to Huang, the new chips, codenamed Blackwell, are much faster than all previously released chips. They will be available for sale this year. It is expected that the new Nvidia chips could cost up to $50,000, which is about double the cost of the previous generation of processors. In the presentation, Huang gave examples of different ways to use the new chips. In particular, they can be used in the development of artificial intelligence systems. He also unveiled new network switches designed for AI development and supercomputing systems that use new chips.
One of the main sources of the B200’s higher performance was the new Tensor Cores and the second generation of the Transformer Engine. The latter has learned to more finely select the required calculation accuracy for certain tasks, which affects both the speed of training and operation of neural networks, and the maximum volume of supported LLMs. Now Nvidia offers AI training in FP8 format, and FP4 is enough to run trained neural networks. But note that Blackwell supports a wide variety of formats, including FP4, FP6, FP8, INT8, BF16, FP16, TF32 and FP64. And in all cases except the last one there is support for sparse computing.
Huang said the new chips will be able to train the latest artificial intelligence models while consuming just 4 megawatts of power for 90 days. Using older chips, the process would require using 8,000 15-megawatt GPUs over the same time period, he said.
