NVIDIA Announces CPU for Massive AI and High-Performance Computing Workloads

• The “Grace” CPU uses energy-efficient Arm cores to deliver a 10x performance boost for systems training giant AI models

• Swiss Supercomputer Center and U.S. Department of Energy’s Los Alamos National Laboratory Build First Supercomputer Powered by NVIDIA CPUs

SANTA CLARA, Calif. — GTC — April 12, 2021 — NVIDIA unveiled its first Arm-based data center CPU processor, delivering 10 times the performance of today’s fastest servers for the most complex AI and high-performance computing workloads.

The result of more than 10,000 engineering years, the NVIDIA Grace™ CPU is designed to meet the computing requirements of the world's most advanced applications - including natural language processing, recommendation systems, and AI supercomputing - that require ultra-fast computing performance and large memory for massive data analysis. It combines energy-efficient Arm CPU cores with an innovative low-power memory subsystem to deliver high performance with high energy efficiency.

“Cutting-edge AI and data science are pushing today’s computer architectures beyond their limits to process data at unimaginable scale,” said Jensen Huang, founder and CEO of NVIDIA. “NVIDIA designed Grace, a CPU designed for large-scale AI and HPC, using IP licensed from Arm. Along with GPUs and DPUs, Grace gives us a third foundational technology for computing and the ability to reimagine the data center to advance AI. NVIDIA is now a three-chip company.”

Grace is a highly specialized processor for workloads such as training next-generation NLP models with more than 1 trillion parameters. When tightly coupled with NVIDIA GPUs, systems powered by Grace CPUs are 10 times faster than today’s most advanced NVIDIA DGX™-based systems running on x86 CPUs.

While the vast majority of data centers are served by existing CPUs, Grace (named after American computer programming pioneer Grace Hopper) will serve a niche market for computing.

The Swiss National Supercomputing Center (CSCS) and the U.S. Department of Energy's Los Alamos National Laboratory were the first to announce plans to build supercomputers powered by Grace to support national scientific research efforts.

NVIDIA launched Grace against the backdrop of exponential growth in data volumes and AI model sizes. Today’s largest AI models contain billions of parameters, and the number of parameters doubles every two and a half months. Training these models requires a new CPU that is tightly coupled to the GPU to eliminate system bottlenecks.

NVIDIA built Grace to take advantage of the tremendous flexibility of the Arm data center architecture. With the launch of new server-class CPUs, NVIDIA is advancing its goal of technological diversity in the fields of AI and HPC, where more choice is key to achieving the innovation needed to solve the world's most pressing problems.

“As the world’s most widely licensed processor architecture, Arm is driving innovation in incredible new ways every day,” said Simon Segars, CEO of Arm. “The launch of NVIDIA’s Grace data center CPU is a clear example of how Arm’s licensing model is enabling an important innovation that will further support the incredible work of AI researchers and scientists around the world.”

Grace's first adopters push the limits of science and AI

CSCS and Los Alamos National Laboratory plan to launch a supercomputer powered by Grace, built by HPE, in 2023.

“The new NVIDIA Grace CPU allows us to merge AI techniques with traditional supercomputing to tackle some of the most difficult problems in computational science,” said Professor Thomas Schulthess, director of CSCS. “We are excited to make this new NVIDIA CPU available to our customers in Switzerland and around the world for processing and analyzing large and complex scientific datasets.”

“This next-generation system will reshape our institution’s computing strategy through an innovative balance of memory bandwidth and capacity,” said Thom Mason, director of Los Alamos National Laboratory. “With NVIDIA’s new Grace CPU, we can perform advanced science with high-fidelity 3D simulations and analysis on larger data sets than ever before.”

Achieving breakthrough performance

Grace’s powerful performance is based on fourth-generation NVIDIA NVLink® interconnect technology, which provides a record-breaking 900 GB/s connection speed between Grace and NVIDIA GPUs, enabling 30 times more total bandwidth than today’s leading servers.

Grace will also leverage the innovative LPDDR5x memory subsystem, which has twice the bandwidth and 10 times the power efficiency of DDR4 memory. In addition, the new architecture provides cache coherence of a single memory address space, combining system and HBM GPU memory to simplify programmability.

Grace will be supported by the NVIDIA HPC software development kit and a full suite of CUDA® and CUDA-X™ libraries to accelerate more than 2,000 GPU applications, enabling faster discovery for scientists and researchers tackling the world’s most critical challenges.