Grace: nvidia’s first arm processor for supercomputer

Grace: Nvidia's first arm processor for supercomputer

Nvidia has introduced the first own arm processor for high performance computing and artificial intelligence. With the name "Grace" Nvidia refers according to CEO Jensen Huang on the US computer pioneer Grace Hopper. In the GPU broken cake, their last name was considered "Hopper" Last as a candidate for the name of the chip. NVIDIA currently ames the CPU developer arm.

Grace should be available from the beginning of 2023 and use in supercomputer systems – such as in the Alps supercomputer of the Swiss National Supercomputing Center. According to Huang, the next generation of NVIDIA computing accelerators should also be used in Alps. For some time, Nvidia’s computing accelerator programming platform CUDA is also usable on servers with arm processors. NVIDIA prepares its own HPC-SDK as well as the CUDA / CUDA X libraries for Grace.

NVIDIA presents Arm-CPU Grace (GDC 2021) (Source: Nvidia)

Alps is built by Hewlett Packard Enterprise on Cray-Ex-Basis and replaces the previous PIZ-DAINT supercomputer.

Grace: Nvidia's first arm processor for supercomputer


Grace with arm neoverse

Nvidia Grace relies on Arm cores from the FUR SERVER family Neoverse, which are to achieve 300 points at the benchmark SPECRATE2017_INT_BASE.

Neoverse cores are also used in Amazon’s Graviton2 processors and Altra-Max processors from Ampere as well as in the art EU processor Rhea.

The architecture of Grace is intended for particularly high transfer rates – these are of the distributed calculation of multiple instances of central importance. The Grace cores can access ECC-protected LPDDR5x memory, which should achieve particularly high transfer rates of via 500 GB / s.

With other graphics processors, Grace communicates with fast, cached interconnects nvidia, according to the fourth NVLink generation CPU-to- (NVIDIA) GPU transfer rates of 900 GB / s possible; from CPU to CPU Still 600 GB / S.

Grace: Nvidia's first arm processor for supercomputer

For developing and arm migrating from HPC applications, NVIDIA offers a HPC Devkit including associated SDK.

CUDA code port

With the help of NVIDIA HPC SDK, scientists and developers will migrate their HPC and KI software to GPU-accelerated ARM systems. To start such migrations before Grace comes to the market, NVIDIA announced together with Amazon EC2 instances in the AWS Cloud with Graviton2 CPU and NVIDIA GPUs.

NVIDIA offers developers an arm HPC Developer Kit, which from an ampere-Altra-CPU with 80 arm-neoverse cores (up to 3.3 GHz), two NVIDIA A100 GPUs (624 FP16 Tflops) and two Bluefield-2 -DPU.

Grace: Nvidia's first arm processor for supercomputer


Like this post? Please share to your friends:
Leave a Reply

;-) :| :x :twisted: :smile: :shock: :sad: :roll: :razz: :oops: :o :mrgreen: :lol: :idea: :grin: :evil: :cry: :cool: :arrow: :???: :?: :!: