NVIDIA Blackwell B200 GPU Announced: Here Are Its Features

NVIDIA, the number one in the GPU world, announced its new chip at an event it organized. This artificial intelligence-focused chip is derived from the H100, which made NVIDIA a multitrillion-dollar company. much better. Even NVIDIA, “The world’s best AI-focused chip” has announced to the whole world that it has produced. So, what do NVIDIA’s new chips promise to artificial intelligence developers?

According to the statement made by NVIDIA Blackwell B200 GPU The chip named is equipped with 208 billion transistors. This allows the chip to produce 20 petaflops of power. According to the company, the new generation transistor engine used in this chip uses a 4-bit system for each neuron. This means that computing, bandwidth and model size are has doubled means.

25 times more efficient than NVIDIA H100!

*Blackwell B200 GPU on the left, H100 on the right.

According to the statement made by NVIDIA, it combines two Blackwell B200 GPUs with a Grace CPU.GB200It created a chipset called “”. According to the company’s statement, this chipset is 30 times more performant than the H100. Moreover, it is just as efficient. According to NVIDIA, the cost and energy consumption are lower than the H100. up to 25 times will decrease.

Let’s say we have 18 trillion parameters (GPT-4 has an artificial intelligence model with approximately 1.7 trillion parameters). According to NVIDIA, 8,000 Hopper GPUs and 15 megawatts of energy consumption are needed to train this artificial intelligence model. The same training, 2,000 Blackwell B200 GPUs and 4 megawatts It can also be given by energy consumption. You can understand how efficient the new chip is from here.

*NVIDIA G200 chipset.

According to NVIDIA, past 16-GPU AI chipsets spent 60 percent of their time communicating, 40 percent was also spent on computing. The new chipset eliminates this too. Company; 576 GPUs connected together in the GB200 1.8 terabytes created a new connection system with bi-directional bandwidth.

*NVIDIA GB200 NVL72.

According to NVIDIA’s opinion, large companies developing artificial intelligence will not purchase a single Blackwell B200 GPU. That being the case, 36 CPUs and 72 GPUsPackages with different options have been created. In this context; NVIDIA GB200 NVL72 is one of them. The self-liquid-cooled package can produce a total of 720 petaflops of power. There are 5,000 separate cables in the system you see above. The total length of these cables is 3.22 kilometers. Developers will be able to use this system for a model with 27 trillion parameters.

NVIDIA already GB200 NVL72 already sold. As expected, the buyers of the new system were Oracle, Microsoft and Google. However, the company has not disclosed how many NVL72s it has sold yet. Meanwhile; NVIDIA will also sell much more advanced versions of this system. Because the company introduced GB200 tens of thousands to be connected designed it.