Artificial intelligence (AI) soon to be more powerful? Promise made by Nvidia, the American semiconductor giant. And as if to walk the talk, its CEO Jensen Huang presented this Monday, March 18, during a conference for developers, his latest innovation: “superchips”.
Named Blackwell, in homage to David Blackwell, the first African-American academic to join the National Academy of Science, these electronic chips would be up to four times more powerful than those of the previous generation used to train existing AI models. In short, real tools to boost AI.
“The rate of progress in computing is insane,” argued Jensen Huang, the group’s boss, assuring that they “need GPUs [Graphics processing unit, des puces disposant d’une capacité de calcul très supérieure à celle des microprocesseurs classiques] bigger”.
Less greedy superchips
Larger of course, but also less energy-consuming. Because AI today still consumes a lot of electricity. Which has also earned him a lot of criticism. First in the sector since the release of its flagship product, the H100 in 2022, Nvidia has worked hard, and the results are there: Nvidia’s “superchips” would be 25 times more efficient in terms of energy efficiency.
Enough to leave its competitors behind, who were already struggling to catch up. Because although they are trying to develop their own chips, the giants Apple, Microsoft and Amazon are currently forced to use Nvidia products in order to keep their promises in terms of AI. The Apple firm has notably agreed to cooperate with Nvidia to equip its virtual reality headset, Vision Pro, marketed since February.
Robots that are more and more human?
And the ambitions of the microprocessor specialist do not stop there. Nvidia also revealed this Monday that it is working on Earth-2 Cloud: a platform for forecasting climate change using simulations generated by supercomputers with AI.
Even more impressive, the group announced that it was close to finalizing the “world’s first human foundation model.” Called the Gr00T project, this innovation will allow robots to understand language, imitate the movements of human beings and interact with the world.
The models “will allow the robot to learn from human demonstrations so that it can help with daily tasks and reproduce human gestures, just by observing us,” the group continues.
A menacing Chinese shadow
The only downside: unlike its rivals Intel, Micron and Texas Instruments, Nvidia does not manufacture its own semiconductors. The company, which in February exceeded the symbolic threshold of 2,000 billion dollars in valuation (a summit that only Microsoft, Apple and the oil company Saudi Aramco have known), in fact uses subcontractors. Among which, the company Taiwan Semiconductor Manufacturing.
An architecture whose foundations could prove fragile. Even more so in the context of the strong geopolitical tensions that Taiwan is experiencing with its Chinese neighbor. Recently, Washington, which is engaged in a trade war with the Middle Kingdom, formally prohibited Nvidia from supplying Chinese companies with its most efficient chips.