NVIDIA continues to expand its investments in artificial intelligence. The company, which made a big impact in the sector with its H100 chip, wants to further consolidate its dominance in the artificial intelligence market. In this context, NVIDIA Blackwell B200, which is called “the world’s most powerful artificial intelligence chip”, was introduced.
NVIDIA Blackwell B200 specs
The NVIDIA Blackwell B200 offers up to 20 petaflops of FP4 computing power thanks to its 208 billion transistors. Even more impressive is the ability to combine two B200 chips with a single Grace CPU to create the GB200 superchip.
NVIDIA claimed that the GB200 superchip can deliver 30x better performance on AI workloads compared to the H100 chip. It can also reduce cost and energy consumption by up to 25 times. According to the company, a single GB200 will be just one card size
NVIDIA stated that a job previously done with 8 thousand Hopper GPUs can be done with only 2 thousand Blackwell GPUs. He also stated that this cluster will provide 75 percent efficiency by consuming only four megawatts of energy. It can be used to train an artificial intelligence model with 1.8 trillion parameters.
NVIDIA attributes the jump in performance to in-house innovations such as the second-generation transformer engine, which can perform 4-bit calculations. It has also developed a new 50 billion transistor NVLink chip for more than 500 GPUs to communicate.
The company said it will bundle multiple GB200 superchips together for large companies. NVIDIA claimed that its technology can now be used to “train models with tens of trillions of parameters.” The company is expected to first offer this chip to companies such as Amazon, Google, Microsoft and Oracle.
{{user}} {{datetime}}
{{text}}