Nvidia has
announced that its Grace CPU and Grace Hopper will build a supercomputer called Venado for Los Alamos National Laboratory (LANL), with peak AI performance expected to exceed 10 ExaFlops. Nvidia said a number of manufacturers, including Atos, Dell Technologies, GIGABYTE, Hewlett Packard Enterprise, Inspur, Lenovo, and Supermicro, will use the Grace series of chips to build next-generation servers.
Venado will be the first system in the U.S. to use Nvidia’s Grace CPU technology, built using the HPE Cray EX supercomputer, a heterogeneous system that uses a mix of Grace CPU Superchips and Grace Hopper chips to accommodate a wide range of emerging programs. Venado is not the only supercomputer built with the Grace series of chips. A supercomputer called Alps at the Swiss National Computing Center (CSCS) also uses a similar design and is said to have a computing power of 20 ExaFlops.
“As supercomputing enters the era of exascale AI, NVIDIA is teaming up with our OEM partners to enable researchers to tackle massive challenges previously out of reach,” said Ian Buck, vice president of Hyperscale and HPC at NVIDIA. “Across climate science, energy research, space exploration, digital biology, quantum computing and more, the NVIDIA Grace CPU Superchip and Grace Hopper Superchip form the foundation of the world’s most advanced platform for HPC and AI.”
Grace CPU Superchip is built on two Grace CPUs, with a total of 144 Arm v9 architecture CPU cores, with a cache capacity of 396MB. The LPDDR5x memory bandwidth with ECC check function reaches 1TB/s, supports PCIe 5.0, DDR5, HBM3, CCIX 2.0, and CXL 2.0, and other features, with a TDP of 500W. The Grace CPU Superchip connects via NVIDIA’s latest NVLink-C2C, which provides 900 GB/s of connection bandwidth to guarantee low latency and consistency between chip-to-chip interconnects and allow connected devices to work on the same memory pool.
Grace Hopper is a combination of NVIDIA’s latest Hopper architecture GPU and Grace CPU, and also uses NVLink-C2C to connect the CPU and GPU. It has 72 Arm v9 architecture CPU cores, and the GPU should be consistent with the H100 computing card, that is, 16896 FP32 CUDA cores, equipped with 600GB of memory.
The Grace CPU Superchip and Grace Hopper can be paired with up to 8
Hopper architecture GPUs in many different configurations. These combinations will use the NVIDIA ConnectX-7 interconnect chip to enable NVLink communication through the built-in PCIe 5.0 switching system, enabling a wider range of systems and applications. According to Nvidia’s schedule, both the Grace CPU Superchip and Grace Hopper will be officially available in the first half of 2023.