Following closely behind Apple, Nvidia has announced a dual processor ‘superchip’ to boost CPU performance in data centre AI applications.

The Grace ‘superchip’ combines two die with the NVLink-C2C interconnect to provide 144 high performance ARM v9 Neoverse cores and 1 Terabyte/s bandwidth to low power DDR5 memory in a single socket.

This is very similar to the Apple M1 Ultra, which combines two of its M1 Max Pro die with a high performance interconnect called UltraFusion:Apple goes to 2.5D for 114bn transistor M1 Ultra

The Grace CPU Superchip will have a total power envelope of 500W and will run all of Nvidia’s computing software stacks, including RTX, HPC, AI and Omniverse. The Grace CPU Superchip along with the ConnectX-7 NIC interface cards offer the flexibility to be configured into servers as standalone CPU-only systems or as GPU-accelerated servers with one, two, four or eight GPUs using the latest Hopper GPU architecture.

Related Grace CPU articles

• Nvidia opens up NVLink for chiplets to take on PCI Express
• Nvidia stakes its claim to the data centre
• 瑞士AI超级计算机使用新的NVIDIA ARM芯片
• NVIDIA揭幕了巨型AI和HPC的CPU

Hopper is Nvidia’s latest GPU architecture which boosts the performance of TensorFlow AI models and adds dynamic programming. This is an algorithmic technique for solving a complex recursive problem by breaking it down into simpler subproblems. By storing the results of subproblems to avoid recomputing them later reduces the time and complexity of problem solving.

Hopper introduces DPX instructions to accelerate dynamic programming algorithms by 40X compared to CPUs and 7X compared to Nvidia’s previous Ampere architecture GPUs.