NVIDIA Grace

Grace is NVIDIA’s first datacenter CPU. All Grace products start with a system-on-chip (SoC) that comprise 72 high-performance Arm v9 cores and feature the NVIDIA-proprietary Scalable Coherency Fabric (SCF) network-on-chip for incredible core-to-core communication, memory bandwidth, and new GPU I/O capabilities. Grace provides a high-performance compute foundation in a low-power system-on-chip.

Built on standards such as Arm SystemReady SR, the Grace CPU is compatible with a wide variety of Arm-compatible operating systems, PCIe and USB peripherals, drivers, and application software already commonplace in existing Arm deployments— whether in the datacenter or the public cloud—including NVIDIA’s CUDA and GPU driver ecosystem.

Grace is available in a variety of platforms for traditional and accelerated compute— including Grace Hopper products like GH200, which integrate a single 72-core Grace CPU with a H100 GPU on a new common memory subsystem to enable the next frontier of accelerated workloads—and the Grace CPU Superchip, which features a dual-CPU configuration with 144 cores, delivering the performance of today’s highest-end conventional 2-socket CPU-based servers while improving datacenter efficiency by 2x.

11/09/23
The Grace Performance Tuning guide provides best practices, software, and hardware configuration suggestions—and an overview of some of the tools available— to maximize performance on all NVIDIA Grace-powered systems.

Grace OS Installation Guides

Grace systems can run a variety of Linux distributions that support the AArch64 architecture. With the proper kernel support and configurations, you can run one of the following Linux distros and take advantage of the advanced Grace features. These guides provide information about how to install a Linux distribution on Grace systems.

12/18/23
Grace systems are composed of multiple hardware components that require support across different software subsystems. This guide provides information about the known patches and configuration settings that are required (or recommended) for interoperability with the Grace platform and might be helpful for customers who want to create their own custom software, for example a Linux kernel, for these systems.
The Grace CPU Benchmarking guide demonstrates how to run popular CPU-intensive applications and benchmarks that are commonly run on X86-64 and other Arm(tm) platforms and includes reference results and procedures that help users and developers understand how to run general purpose CPU workloads to their full potential on all Grace-powered platforms.
NVIDIA Grace CPU SuperChip-based MGX servers are shipping now. These servers are suitable for most applications where high-performance x86-64 or Arm systems are commonly deployed and run conventional operating systems and software. The servers can be used as standalone compute servers, clustered for HPC use cases, and can accommodate acceleration with the addition of optional PCIe GPUs. The guide covers the unboxing and basic operating system installation and configuration of a typical platform.
NVIDIA GH200 SuperChip-based MGX servers are shipping now. These servers combine a Hopper GPU with a 72-core Grace CPU and memory on a module and are designed for accelerated applications that are tightly coupled with the underlying CPU and memory platform. This guide covers the unboxing and basic operating system installation, driver and runtime setup, and configuration of a typical platform.
This application note provides NVIDIA GH200 benchmark data in comparison to the NVIDIA® DGX™ H100 platform.