After the introduction of the Denver CPU project and the Tegra K1 mobile processors, Nvidia’s CPU activity has been unheard of for several years. But now the company has returned to the CPU field with a new project called Nvidia Grace, and wants to build a chip based on a logo architecture for data centers focused on artificial intelligence.
Nvidia announced a $ 40 billion deal to buy the logo last September and is now working to obtain the necessary licenses for the deal from legal authorities. The company claims that with this contract, it wants to prepare the logo for the age of artificial intelligence. Now the design of this chip can also be a step in the right direction That should be the goal.
The new Grace processor is named after American computer expert Grace Hopper. Launched in 2023, the chip will be 10 times more powerful than the world’s fastest servers in running the most sophisticated AI processes. With this in mind, Grace could be a viable option for research center supercomputers.
According to Nvidia’s roadmap, the image of which was released at the GTC 2021 conference, the company will introduce the second generation of Grace in 2025. Nvidia has not yet released specific information about the technical specifications of the grease, but we know that the chip uses the fourth generation NVLink, which connects the CPU and GPU at an astonishing speed of 900 Gbps.
This figure is even higher than the CPU memory bandwidth, so the GPU will have a special link with the CPU to access the system memory with maximum bandwidth. This method actually allows the whole system to use a common memory address space.
Nvidia will use the next generation of Neoverse cores in the Grace chip, which the company claims can score above 300 in the SPECrate2017_int_base benchmark. Nvidia has not yet released more information on how to configure or optimize the processor’s neural network.