Nvidia has unveiled its latest artificial intelligence chips, which the company claims are far more powerful than its existing market-leading hardware, as it sets its sights on extending its domination of the burgeoning industry.
Chief executive Jensen Huang on Monday said Nvidia’s Blackwell graphics processing units will massively increase the computing power driving large language models. The Blackwell GPU has 208bn transistors, compared with last year’s H100, which has 80bn, a measure of its increased power.
Huang said the chip was twice as powerful when it came to training AI models as its current generation of GPUs, and had five times their capability when it came to “inference” — the speed at which AI models such as ChatGPT can respond to queries.
“The inference capability of Blackwell is off the charts,” he said, adding that there was “unbelievable excitement” about it.
Huang said Nvidia’s GB200 “superchip” would combine two of the Blackwell GPUs with the company’s existing “Grace” central processing unit, which has powered the growth of generative AI.
Huang’s keynote address — delivered in front of a packed audience of developers, investors, media and others at the SAP Center in San Jose, California, where Nvidia is holding its annual developer conference — underlined the buzz the company has attracted as the excitement about AI has grown. In two years, Nvidia has managed to turn the debut of new GPU chips into an event equalling the hype of consumer products launches from the likes of Apple, a phenomenon Huang himself noted on stage.
When Nvidia first went to the market around two years ago with its “Hopper” chips, it started with just two potential customers, Huang said. “We have more now,” he quipped.
That the Nvidia executive was going to announce the Blackwell chips during his presentation was one of the worst kept secrets in Silicon Valley, and the presentation did little to jolt investors. Nvidia’s market capitalisation has grown to $2.2tn, overtaking Google and Amazon to become the third most-valuable company in the world after Microsoft and Apple, but its shares were slightly lower in after-hours trading following the event.
The new partnerships, customers and software tools Huang revealed during his two-hour speech reflected Nvidia’s core contention: that AI will have a transformative effect on every part of the global economy.
All the big cloud service providers, which include Google, Amazon and Microsoft, were “lined up as customers” for what he predicted would be “the most successful product launch in our history”.
Nvidia’s chips have become the commodity that has underpinned the generative AI revolution, and its rocketing stock has fuelled share price rallies across the globe.
Companies including OpenAI, which is backed by Microsoft, and other entities have spent billions of dollars to acquire Nvidia’s chips, which offer the computing power behind generative AI technology that can produce humanlike text, video and code within seconds.
At the same time, Google, Amazon and Microsoft are racing to develop their own generative AI chips, in an attempt to reduce their dependence on Nvidia and to lock customers more closely into their own hardware and software systems.
Nvidia’s new focus on AI inference addresses Wall Street concerns that the huge demand for its chips may start to tail off as companies such as Intel and AMD seek to compete in the space and customers shift away from training large language models to implementing them.
To help entrench itself as the primary AI chip supplier, Nvidia is also building out its software offering with tools to help build AI applications that run on the chips.
The company is releasing a new layer to its Cuda software called “NIM microservices”, which businesses can use to deploy a range of AI models that have been optimised to run on Nvidia’s chips.
“We are effectively an AI foundry . . . We will do for you and the industry on AI what TSMC does for us building chips,” he said, pointing to new partnerships with a number of companies including SAP, Snowflake, and NetApp. Taiwan’s TSMC is the world’s leading chip manufacturer and a critical component in the global supply chain.
Whether Nvidia can continue its trajectory depends in part on the returns its customers start to see on their initial investment. With investors rushing into Nvidia’s stock, the pressure is on for the company to maintain its dizzying growth.
The ultimate size of this future generative AI market remains unclear. Huang has predicted that the total value of all the technology in data centres will reach $2tn in the next four to five years.
Huang also used the event to promote Nvidia’s Omniverse software platform, which can create “digital twins” of objects. In the future, “we’re going to manufacture everything digitally first, and then we’ll manufacture it physically”, he said.
New Nvidia AI models and software would help with the development of humanoid robots that look and move like people, which Huang said would be “the next generation of robotics”. Towards the end of the event, Huang invited two small robots on to the stage, one of which at a certain point seemed to ignore his instructions to move closer to him.
The market’s muted reaction may point to the frenzy around Nvidia “hitting its peak”, Daniel Newman, chief executive of The Futurum Group, told the Financial Times. “The market is very forward-looking. It already sees massive market share, the dominant position with the hyperscalers, the genuine market lead it has with the stickiness of Cuda [software].”
The arrival of competing chips coincides with the question of how long Nvidia can sustain double-digit growth quarter after quarter, Newman added. “There’s only really one way for Nvidia to go: we need to see how big this market really is.”
Read the full article here