During Computex 2022, Nvidia announced the upcoming release of its first Nvidia Grace CPU-powered system reference designs. Upon launch, Nvidia’s first CPU will help usher in the next generation of high-performance computing (HPC), enabling tasks like complex artificial intelligence, cloud gaming, and data analytics.
The upcoming Nvidia Grace CPU Superchip and the Nvidia Grace Hopper Superchip will find their way into server models from some of the most well-known manufacturers, including Asus, Gigabyte, and QCT. Along with x86 and other Arm-based servers, Nvidia’s chips will bring new levels of performance to data centers. Both the CPU and GPU were initially revealed earlier this year, but now, new details have emerged along with an approximate release date.
Although Nvidia is primarily known for making some of the best graphics cards, the Grace CPU Superchip has the potential to tackle all kinds of HPC tasks, ranging from complex AI to cloud-based gaming. Nvidia teased that the Grace Superchip will come with two processor chips connected via Nvidia’s NVLink-C2C interconnect technology.
Together, the chips will offer up to 144 high-performance Arm V9 cores with Scalable Vector Extensions, as well as an impressive 1TB/s memory subsystem. According to Nvidia, its new design will double the memory bandwidth and power efficiency of current generation server processors. Some of the use cases for the new CPU that Nvidia lists include data analytics, cloud gaming, digital twins, and hyperscale computing applications.
Launching alongside nvidia Grace is the Nvidia Grace Hopper Superchip, and while strikingly similar by name – the “Hopper” gives it away – this isn’t just a CPU. Nvidia Grace Hopper pairs an Nvidia Hopper graphics card with an Nvidia Grace processor, once again using the same NVLink-C2C technology.
The combination of the two has a massive effect on data transfer speed, making it up to 15 times faster than traditional CPUs. Both chips are impressive, but the Grace and Grace Hopper combo should be able to tackle just about any task, including giant-scale AI applications.
Nvidia’s new server design portfolio offers single motherboard systems with up to four-way configurations available. These layouts can be further customized based on individual needs to fit specific workloads. To that end, Nvidia lists a few systems.
The Nvidia HGX Grace Hopper system for AI training, inference, and HPC comes with Nvidia’s Grace Hopper Superchip and BlueField-3 data processing units (DPUs). There is also a CPU-only alternative that combines the Grace CPU Superchip with BlueField-3.
Nvidia’s OVX systems are aimed at digital twins and collaboration workloads and come with a yet-to-be-revealed Grace CPU chip, BlueField-3 and Nvidia GPUs. Lastly, the Nvidia CGX system is made for cloud gaming and graphics. It pairs the Grace CPU Superchip with BlueField-3 and Nvidia’s A16 GPUs.
Nvidia’s new lineup of HPC processors and graphics cards will launch in the first half of 2023. The company has teased that dozens of new server models from its partners will be available by then.