Nvidia Blackwell B200 and GB200 are the most powerful AI GPUs yet

Nvidia has launched Blackwell B200 and GB200, the two most powerful AI GPUs at the GPU Technology Conference (GTC) 2024. The chipsets are named after David Harold Blackwell, an American statistician and mathematician who made several contributions to game theory, as well as probability and information theory. He was also the first African American inducted into the National Academy of Sciences.

Nvidia Blackwell B200 and GB200: What you need to know

According to CEO Jensen Huang who unveiled the GPUs, the B200 GPU offers speeds of up to 20 petaflops, being armed with as many as 208 billion transistors. A petaflop is a unit of measurement that measures the computing power of a processor and the number of floating-point calculations a computer can perform in seconds.

ALSO READ: Nvidia’s new AI chatbot runs on your PC

The higher the number of petaflops, the faster your computer is. A GPU with a speed of 20 petaflops can perform as many as 20,000 thousand trillion operations per second. If that’s not good enough, you have the Blackwell GB200, which combines two Blackwell B200 GPUs with the company’s Grace central processing unit offering as much as 30 times the performance of both.

What will the Nvidia Blackwell B200 and GB200 power?

Earlier this year, Nvidia overtook Amazon and Alphabet to become the third most valuable company after Apple and Microsoft. The launch of the two GPUs is in line with the company’s focus to be the driver of the world’s AI infrastructure.

The Blackwell GB200 is expected to be able to train LLMs seven times faster than Nividia’s H100 that used to be one of the most powerful chips for AI. What’s more, speed will not come at the expense of more power. For instance, to train an LLM with up to 1.8 trillion parameters, it would have earlier taken up to 8,000 H100 GPUs and consumed up to 15MW of power. Now, it can be accomplished using only 2,000 Blackwell GPU, consuming as little as 4MW of power.

Now, Nvidia is packing up to 36 CPUs and 72 GPUs in larger, liquid-cooled racks such as the GB200 NVL72. The liquid-cooled rack gets up to 1,440 petaflops of inference and has nearly two miles of cables packed inside. Each tray in the rack gets two GB200 chips. This will support up to a 27 trillion parameter model. GPT-4, which is one of the largest and most powerful LLMs there is today, is an (approximately) 1.7 trillion parameter model.

ALSO READ: ChatGPT is getting human-like memory; will soon remember you and things you like

With generative AI set to play a pivotal role across industries and workstreams, applications of the Nvidia Blackwell chipsets are infinite. Whether its powering heavy industries or running efficient supply chains, the Nvidia B200 and GB200 will make processing faster and more efficient.

Nvidia Black B200 and GB200: What these chipsets mean for you

While Blackwell B200 and GB200 are set to power the engines that drive the global economy, they will also have, how shall we put it, everyday applications too. The Blackwell GPU architecture will also likely power a few chips in Nvidia’s upcoming RTX 50 series of desktop graphics cards. So, the new GPUs will be significantly more powerful than the RTX 4090, Nvidia’s current flagship models for laptops and desktops, and will also have generative AI capabilities on board.

So your machine will be able perform tasks such as generation of text, images or videos, or even edit them, as well as debug code, all in real time. What’s more, it will also do these things faster and more accurately than ever before. This means that not only will your get done quicker, but also require little to no intervention in the near future, putting aside basic quality checks.