Nvidia’s new Grace CPU puts it in direct competition for data-center business with incumbent rivals Intel and AMD. Together, AMD and Intel form a powerful duopoly that has captured over 90% of the server chip market, according to Jefferies Research.
Nvidia’s server processor revolves around a core designed by Arm, which has attracted more interest as big tech companies such as Amazon design their own data-center processors. Nvidia unsuccessfully attempted to buy Arm but secured a 20-year license for its tech after abandoning the deal earlier this year.
Much like Apple’s recently unveiled M1 Ultra processor, Nvidia’s Grace CPU Superchip for servers will feature two CPUs joined by a custom-designed interconnect intended to deliver high-speed data transfers with low latency. Based on the Arm v9 tech, the Grace CPU will have 144 high-performance cores — Nvidia claims it will double the performance and energy efficiency of existing “leading” server chips.
In materials distributed ahead of its semi-annual GTC conference, the company said it plans to bring the interconnect technology, called NVLink, to all of its future chips.
The new Grace processor is designed for AI, high-performance computing and hyperscale data center applications. The company plans to release the processor and its previously announced combination CPU-GPU chip called the Grace Hopper Superchip in the first half of 2023.
“Grace is the ideal CPU for the world’s AI infrastructures,” Nvidia CEO Jensen Huang said in a statement. Huang was scheduled to give a keynote address at GTC Tuesday outlining additional details about the new chip.
Roughly two years after the launch of its prior generation Ampere graphics chip architecture, Nvidia said Tuesday that it planned to sell its H100 next-generation server graphics processors based on the new Hopper architecture in the third quarter of this year. The server chips will feature 80 billion transistors manufactured on TSMC’s four-nanometer fabrication process.
Nvidia said the Hopper-based chip includes a number of enhancements, including a natural language processing engine, increased security features that allow for confidential computing and the ability to segment each GPU into smaller parts to handle different types of jobs.
In addition to the new chip and architecture, Nvidia also plans to launch a new generation of its DGX server systems that will feature the new technology. The company said AWS, Google and Microsoft all plan to offer H100-based instances.