NVIDIA: A100 and H100 liquid-cooled accelerators and partnership with Taiwanese for Grace CPUs | Computex 2022

NVIDIA announced at Computex 2022 several partnership with the main companies of Taiwan in view of its entry into the world of CPU per i server intended for the HPC and AI sector. Dozens of servers designed by ASUS, Foxconn Industrial Internet, GIGABYTE, QCT, Supermicro e Wiwynn will arrive on the market in the first half of 2023. Project-based systems Grace CPU Superchip e Grace Hopper Superchip they will be based on four new customizable designs to meet whatever needs the market desires.

We remember that Grace CPU Superchip is a processor that integrates two Arm v9 CPUs with SVE extensions, each equipped with 72 cores for a total of 144 units, on a single module. The two CPUs are interconnected via ultra-high speed NVLink-C2C and connected to a memory subsystem that ensures a bandwidth of 1 TB / s. Further details can be found here.

Grace Hopper Superchip combines a Grace CPU to a Hopper GPU on a single moduleagain taking advantage of the NVLink-C2C interconnection capable of transferring data from the CPU to the GPU with a speed of 15 times faster than traditional CPUs.

NVIDIA also announced the arrival of A100 80GB and H100 accelerators in PCI Express format in a version ready for the liquid cooling. More compact than air-cooled solutions, the new solutions will arrive respectively in Q3 (A100) and early next year (H100) to improve energy efficiency and server density.

The cards come with a waterblock that completely covers the card but which, being single slot, allows you to integrate multiple accelerators within a chassis. The technical specifications of the cards, cooling aside, do not change. The new solutions therefore aim to improve energy efficiency and density rather than performance.

NVIDIA partnered with Equinix to qualify the liquid-cooled version of A100 and evaluate the reduction in overall energy consumption: a configuration made up of 2000 servers (4000 A100 accelerators) saw its total energy requirement reduce by 28%. According to NVIDIA, this stems from a combination of factors, including running the throttle in cooler temperatures to the less energy-intensive cooling system than the air-powered one.

In recent months the company has also worked on liquid design for HGX systems equipped with accelerators in SMX format. A liquid-cooled HGX A100 is already shipping and a liquid-cooled HGX H100 is expected to debut in the fourth quarter.