Nvidia’s new generative AI supercomputer may be one of the most powerful ever made

Nvidia—which already dominates AI computing and is closing in on a $1 trillion valuation—has spat out another massive advance in raw processing power.

On May 28, the chipmaker announced the DGX GH200 AI supercomputer, which runs on 256 GH200 Grace Hopper Superchips. According to an Nvidia press release, the new system uses NVLink interconnect and switch technology to combine the chips into a single functional GPU with “1 exaflop of performance and 144 terabytes of shared memory.”

Nvidia CEO Jensen Huang told the Computex conference in Taiwan that the DGX GH200 AI system combines “Nvidia’s most advanced accelerated computing and networking technologies to expand the frontier of AI.” As Engadget noted, the only other exaflop-scale computer in operation today is Oak Ridge National Laboratory’s Frontier, which is rated at 1.1 exaflops.

Barron’s reported that the company says Google, Meta, and Microsoft will be among the first major companies with access to DGX GH200 AI, which is scheduled to launch by the end of 2023. Huang also announced other generative AI capabilities, particularly in gaming, where Nvidia hopes its Avatar Cloud Engine for Games will enable developers to build NPCs that can interact with players in a more realistic fashion.

Nvidia’s rivals for the generative AI computing throne are mainly Intel and AMD—although Huang raised the point in Taiwan that Intel’s ongoing pivot to operating as a foundry for other chipmakers could lead to partnerships in the future. (While Intel’s stock price has plummeted over the past few years and the company’s foundry ambitions have run into serious obstacles, some investors are betting it’s nearly at the end of the tunnel.) AMD CEO Lisa Su has positioned AMD as Nvidia’s main contender and is betting on an aggressive strategy to catch up to it in the AI computing market.