(Credit: Nvidia)
To meet the demands for AI computing, Nvidia is introducing the Blackwell architecture to produce its fastest GPU chips yet.
Depending on the task, Blackwell GPUs promise to deliver a seven to 30 times performance improvement over Nvidia’s H100 product, which tech companies across the globe have been buying to train AI programs such as ChatGPT.
“Blackwell GPUs are the engine to power this new industrial revolution,” said Nvidia CEO Jensen Huang while introducing the chip at the company’s GTC event in San Jose.
(Credit: Nvidia)Despite the huge performance gains, the new chip uses up to 25 times less energy, the company says. Blackwell is built with an improved version of TSMC’s 4-nanometer process, which Nvidia previously used to produce the H100, first introduced in 2022. The big difference is that Blackwell contains a whopping 208 billion transistors, up from 80 billion in the H100.
The company pulled this off by combining two large chip dies, which communicate over a 10TB/s interlink, to create a unified GPU.
(Credit: Nvidia)Nvidia plans on selling the technology via the GB200 Superchip, which will contain two Blackwell GPUs and an integrated Grace GPU processor. Blackwell will also be sold through a server product called the GB200 NVL72, which will contain 72 Blackwell GPUs all connected over the company’s upgraded NVLink technology to deliver a “1.8TB/s bidirectional throughput.”
Whether Blackwell will one day end up in gaming GPU was left unclear. But the technology will power the latest craze in tech: generative AI. Nvidia has been reporting record earnings on the tech world’s growing appetite for GPUs to fuel AI development. There’s no word on Blackwell’s cost, but Nvidia says Microsoft, Meta, OpenAI, Google, and Amazon all plan on adopting the technology. The first Blackwell GPUs are slated to launch later this year.