Nvida unveils Vera Rubin, Blackwell Ultra, its two new AI ‘superchips

Nvidia

Nvidia announced a major leap in AI computing at its annual GTC conference, unveiling new chips designed to enhance artificial intelligence model development and deployment.

CEO Jensen Huang introduced Blackwell Ultra, an upgraded version of the company’s current Blackwell AI chips, set to ship in the latter half of 2024. Nvidia also revealed Vera Rubin, its next-generation GPU architecture, expected to debut in 2026.

For the first time, Nvidia is introducing a custom-designed CPU, Vera, based on a proprietary core named Olympus. Previously, Nvidia relied on off-the-shelf Arm CPU designs, but the new in-house chip promises double the speed of last year’s Grace Blackwell CPU.

The Rubin GPU, when paired with Vera, will deliver 50 petaflops for AI inference—more than double the 20 petaflops offered by Blackwell. It will also support up to 288GB of high-speed memory, a critical specification for AI developers. Nvidia is redefining how it categorises GPUs, treating its multi-die configurations as separate units rather than a single chip. By 2027, it plans to launch “Rubin Next”, a chip combining four dies, effectively quadrupling Rubin’s performance, revealed Huang.

The Blackwell Ultra series will deliver higher AI processing speeds, generating more content per second. It will be available in different configurations, including GB300, which pairs two GPUs with an Nvidia Arm CPU, and B300, which features only the GPU.

Furthermore, Nvidia will offer versions with eight GPUs per server blade and rack-mounted systems with 72 GPUs.

Nvidia says cloud providers using Blackwell Ultra can generate 50 times more revenue than those relying on older Hopper-generation chips. The world’s top cloud companies have already deployed three times as many Blackwell chips as Hopper, signaling strong industry demand for Nvidia’s AI solutions, said Huang.