Nvidia pronounces GB200 Blackwell AI chip, launching later this 12 months

Nvidia on Monday introduced a brand new technology of synthetic intelligence chips and software program for working synthetic intelligence fashions. The announcement, made throughout Nvidia’s developer’s convention in San Jose, comes because the chipmaker seeks to solidify its place because the go-to provider for AI firms.

Nvidia’s share worth is up five-fold and whole gross sales have greater than tripled since OpenAI’s ChatGPT kicked off the AI growth in late 2022. Nvidia’s high-end server GPUs are important for coaching and deploying giant AI fashions. Firms like Microsoft and Meta have spent billions of {dollars} shopping for the chips.

The brand new technology of AI graphics processors is known as Blackwell. The primary Blackwell chip is known as the GB200 and can ship later this 12 months. Nvidia is attractive its clients with extra highly effective chips to spur new orders. Firms and software program makers, for instance, are nonetheless scrambling to get their palms on the present technology of “Hopper” H100s and related chips.

“Hopper is improbable, however we’d like larger GPUs,” Nvidia CEO Jensen Huang mentioned on Monday on the firm’s developer convention in California.

Nvidia shares fell greater than 1% in prolonged buying and selling on Monday.

The corporate additionally launched revenue-generating software program referred to as NIM that may make it simpler to deploy AI, giving clients another excuse to stay with Nvidia chips over a rising discipline of rivals.

Nvidia executives say that the corporate is changing into much less of a mercenary chip supplier and extra of a platform supplier, like Microsoft or Apple, on which different firms can construct software program.

“Blackwell’s not a chip, it is the identify of a platform,” Huang mentioned.

“The sellable industrial product was the GPU and the software program was all to assist individuals use the GPU in numerous methods,” mentioned Nvidia enterprise VP Manuvir Das in an interview. “After all, we nonetheless try this. However what’s actually modified is, we actually have a industrial software program enterprise now.”

Das mentioned Nvidia’s new software program will make it simpler to run packages on any of Nvidia’s GPUs, even older ones that is perhaps higher suited to deploying however not constructing AI.

“If you happen to’re a developer, you have obtained an attention-grabbing mannequin you need individuals to undertake, when you put it in a NIM, we’ll guarantee that it is runnable on all our GPUs, so that you attain lots of people,” Das mentioned.

Each two years Nvidia updates its GPU structure, unlocking a giant bounce in efficiency. Lots of the AI fashions launched over the previous 12 months have been educated on the corporate’s Hopper structure — utilized by chips such because the H100 — which was introduced in 2022.

Nvidia says Blackwell-based processors, just like the GB200, supply an enormous efficiency improve for AI firms, with 20 petaflops in AI efficiency versus 4 petaflops for the H100. The extra processing energy will allow AI firms to coach larger and extra intricate fashions, Nvidia mentioned.

The chip consists of what Nvidia calls a “transformer engine particularly constructed to run transformers-based AI, one of many core applied sciences underpinning ChatGPT.

The Blackwell GPU is giant and combines two individually manufactured dies into one chip manufactured by TSMC. It would even be out there as a complete server referred to as the GB200 NVLink 2, combining 72 Blackwell GPUs and different Nvidia elements designed to coach AI fashions.

Amazon, Google, Microsoft, and Oracle will promote entry to the GB200 by means of cloud companies. The GB200 pairs two B200 Blackwell GPUs with one Arm-based Grace CPU. Nvidia mentioned Amazon Internet Companies would construct a server cluster with 20,000 GB200 chips.

Nvidia mentioned that the system can deploy a 27-trillion-parameter mannequin. That is a lot bigger than even the most important fashions, similar to GPT-4, which reportedly has 1.7 trillion parameters. Many synthetic intelligence researchers consider larger fashions with extra parameters and knowledge might unlock new capabilities.

Nvidia did not present a value for the brand new GB200 or the methods it is utilized in. Nvidia’s Hopper-based H100 prices between $25,000 and $40,000 per chip, with complete methods that value as a lot as $200,000, in accordance with analyst estimates.

Nvidia additionally introduced it is including a brand new product named NIM, which stands for Nvidia Inference Microservice, to its Nvidia enterprise software program subscription.

NIM makes it simpler to make use of older Nvidia GPUs for inference, or the method of working AI software program, and can enable firms to proceed to make use of the tons of of hundreds of thousands of Nvidia GPUs they already personal. Inference requires much less computational energy than the preliminary coaching of a brand new AI mannequin. NIM allows firms that need to run their very own AI fashions, as a substitute of shopping for entry to AI outcomes as a service from firms like OpenAI.

The technique is to get clients who purchase Nvidia-based servers to join Nvidia enterprise, which prices $4,500 per GPU per 12 months for a license.

Nvidia will work with AI firms like Microsoft or Hugging Face to make sure their AI fashions are tuned to run on all appropriate Nvidia chips. Then, utilizing a NIM, builders can effectively run the mannequin on their very own servers or cloud-based Nvidia servers with no prolonged configuration course of.

“In my code, the place I used to be calling into OpenAI, I’ll change one line of code to level it to this NIM that I obtained from Nvidia as a substitute,” Das mentioned.

Nvidia says the software program may also assist AI run on GPU-equipped laptops, as a substitute of on servers within the cloud.