New processor, supercomputer and PC on ARM: new NVIDIA products in 2021
At GTC 2021, NVIDIA unveiled its proprietary Grace processor based on ARM architecture. High performance in AI operations and energy efficiency are the key features of the new product. In addition, the company spoke about the upcoming supercomputer based on the new platform and promised to create a new class of desktops with ARM processors.
As part of the Grace platform, ARM cores are used, as well as an energy-efficient memory subsystem based on LPDDR5X modules, which, according to the developers, are capable of providing a 10-fold increase in speed compared to DDR4. To simplify programming, the Grace architecture uses an algorithm to combine system memory and GPU memory HBM. In addition, the vendor intends to create a new class of PCs based on ARM processors with NVIDIA RTX graphics accelerators.
The ARM processor Grace was designed to provide the most efficient and fast execution of the most complex computations that require processing huge amounts of data and a large amount of memory. Thus, the Swiss National Supercomputing Center (CSCS) and Los Alamos National Laboratory plan to commission supercomputers based on new chips in the performance of Hewlett Packard Enterprise in 2023. The new infrastructure will replace the existing Piz Daint supercomputer at CSCS and will serve as a universal system open to the wider research community in Switzerland and around the world.
NVIDIA also unveiled the NVIDIA Arm HPC Development Kit. It includes an Ampere Altra processor with 80 ARM Neoverse cores at up to 3.3 GHz, two NVIDIA A100 GPUs with FP16 deep learning performance of 312 teraflops each, and two NVIDIA BlueField-2 DPUs that accelerate network, storage, and safety. Developers can already apply for early access, and the kit itself will be available in the third quarter of 2021.
Systems that will be built on the Grace platform will receive support for the fourth generation of the NVIDIA NVLink bus, capable of providing a bandwidth of 900 GB/s, exceeding the performance of current server solutions by 30 times. The company also said it is actively working on deploying an ARM platform for cloud-based HPC and launching Amazon EC2 instances. The potential performance of the AI computing of the novelty is estimated at 2.8 exaflops.