NVIDIA showcases Grace reference designs with its first Arm CPU at Computex 2022
NVIDIA went mum on new GeForce announcements at Computex 2022, but it did showcase some new reference designs powered by its first Arm CPU. The first wave of systems with NVIDIA Grace CPU Superchip and Grace Hopper Superchip will ship in H1 2023. These new reference designs will be used to power a new class of data center systems for high-performance computing (HPC), cloud graphics and gaming, and more.
According to NVIDIA, these new systems will co-exist alongside its existing lineup of server systems. This means the company will continue to support x86 processors from AMD and Intel, at least for the foreseeable future. The NVIDIA Grace CPU Superchip is the CPU-only Arm chip that combines two processor chips connected through Nvidia’s NVLink-C2C interconnect technology. There’s also the Grace Hopper Superchip which combines a Hopper GPU and the Grace CPU on a single board using the same interconnect technology.
NVIDIA says the Grace CPU Superchip will be the fastest processor on the market when it ships in early 2023. With two Grace CPUs in tow, the NVIDIA Grace CPU Superchip will offer up to 144 high-performance Arm V9 cores and an impressive 1TB/s memory subsystem. It has partnered with dozens of OEMs that are already working on building various systems based on one of NVIDIA’s four reference designs. These servers will be made available in 1U and 2U form factors with air and liquid cooling systems.
Reference designs with NVIDIA Grace CPU Superchip
As you can see, the CGX systems for cloud graphics and gaming applications will use the dual-CPU Grace Superchip paired with the company’s A16 GPUs based on its Ampere architecture. Similarly, the OVX servers designed for digital twin and omniverse applications also use the dual-CPU Grace Superchip. However, these systems appear to be more flexible with different GPU pairings. Lastly, the HGX platform will be available in two flavors, one of which will use the dual-CPU Grace Superchip while the other will embrace the Grace CPU + Hopper GPU Superchip combination for AI training and other HPC tasks.
The dual-CPU HGX Grace CPU Supechip will support up to 1TB of LPDDR5x memory with up to 1TB/s of memory bandwidth. The HGX Grace Hopper Superchip, on the other hand, will provide 512 GB of LPDDR5x memory, 80GB of HBM3, and up to a combined 3.5 TB/s of memory throughput. You can learn more about these two HGX high-density server reference models below:
As we mentioned earlier, NVIDIA will continue to support x86 processors from AMD and Intel systems well into the future. The reference CGX, OVX, and HGX systems will ship in the first half of 2023.