Nvidia CEO Jensen Huang compared the size of the new Blackwell chip to the current Hopper H100 chip at the company’s developer conference in San Jose, California.
Nvidia
The new generation AI graphics processor is named Blackwell. The first Blackwell chip is called GB200 and is expected to ship later this year. Nvidia is luring customers with more powerful chips and driving new orders. For example, companies and software makers are still scrambling to get their hands on the current generation “Hopper” H100 and similar chips.
“Hopper is great, but it requires a bigger GPU,” Nvidia CEO Jensen Huang said Monday at the company’s developer conference in California.
The company also introduced revenue-generating software called NIM to ease the adoption of AI. This gives customers another reason to stick with his Nvidia chips instead of a growing field of competitors.
Nvidia executives say the company is becoming less of a mercenary chip provider and more of a platform provider on which other companies can build software, much like Microsoft or Apple.
“Blackwell is the name of the platform, not the chip,” Huang said.
“The commercial product that sold was the GPU, and the software was all about enabling people to use the GPU in different ways,” Manuvir Das, Nvidia enterprise VP, said in an interview. “Of course, we’re still doing that. But what’s really changed is that we’re actually in the commercial software business.”
Das said Nvidia’s new software makes it easier to run programs on any of Nvidia’s GPUs, even older ones that may not be suitable for building AI.
“If you’re a developer, you probably have an interesting model that you want people to adopt. If you put it in NIM, you make sure it runs on all GPUs and you get it to a lot of people. ” said Das.
Nvidia’s GB200 Grace Blackwell superchip. Powered by two B200 graphics processors and one Arm-based central processor.
Nvidia updates its GPU architecture every two years, significantly improving performance. Many of the AI models released in the past year were trained on the company’s Hopper architecture, which was introduced in 2022 and is used in chips such as the H100.
Nvidia says Blackwell-based processors like the GB200 offer AI companies a significant performance upgrade, with AI performance of 20 petaflops compared to 4 petaflops for the H100. Nvidia says the additional processing power will allow AI companies to train larger and more complex models.
The chip includes what Nvidia calls a “trans engine built specifically to run transformer-based AI, which is one of the core technologies behind ChatGPT.”
Blackwell GPUs are large, combining two separately manufactured dies into a single chip made by TSMC. It will also be available as a whole server called GB200 NVLink 2, which combines 72 Blackwell GPUs and other his Nvidia parts designed to train AI models.
Amazon, Google, Microsoft, and Oracle sell access to GB200 through their cloud services. The GB200 combines two B200 Blackwell GPUs and one Arm-based Grace CPU. Nvidia said Amazon Web Services will build server clusters with 20,000GB200 chips.
Nvidia said the system can deploy 27 trillion parameter models. This is much larger than the largest models such as GPT-4, which is said to have 1.7 trillion parameters. Many artificial intelligence researchers believe that larger models with more parameters and data can unlock new capabilities.
Nvidia has not disclosed the cost of the new GB200 or the systems in which it is used. Analysts estimate that Nvidia’s Hopper-based H100 will cost between $25,000 and $40,000 per chip, with the entire system costing as much as $200,000.
Nvidia also plans to sell the B200 graphics processor as part of a complete system that occupies an entire server rack.
Nvidia also announced that it is adding a new product called NIM to Nvidia enterprise software subscriptions.
NIM makes it easier to use older Nvidia GPUs for inference and AI software execution processes, allowing enterprises to continue using the hundreds of millions of Nvidia GPUs they already own. Inference requires less computational power than the initial training of a new AI model. NIM enables companies that want to run their own AI models, rather than purchasing access to AI results as a service from companies like OpenAI.
The strategy is to have customers who buy Nvidia-based servers sign up for Nvidia Enterprise, which costs $4,500 per GPU per year to license.
Nvidia works with AI companies like Microsoft and Hugging Face to ensure that AI models run on all compatible Nvidia chips. NIM then allows a developer to efficiently run the model on her own server or her cloud-based Nvidia server without a time-consuming configuration process.
“In my code that was calling OpenAI, I’m going to replace that one line of code and point it to this NIM that I got from Nvidia instead,” Das said.
Nvidia says the software also helps run AI on GPU-equipped laptops rather than servers in the cloud.
