Skip to content

NVIDIA Grace

The NVIDIA Grace CPU is a high-performance ARM-based processor designed for data centers and AI workloads. Named after computer scientist Grace Hopper, it was announced in 2021 and represents NVIDIA’s entry into ARM-based CPUs for HPC, AI, and cloud computing.

Key Specifications

  • Architecture: ARM Neoverse cores (custom NVIDIA variant)
  • Core Count: 72–144 cores depending on configuration
  • Clock Speed: Up to 3.0 GHz
  • Memory: DDR5 or HBM (High Bandwidth Memory) depending on system design
  • Interconnect: NVLink for high-speed GPU-CPU connectivity
  • Focus: Optimized for AI, HPC, and high-throughput workloads

Architectural Features

  1. High Core Count
    NVIDIA Grace CPUs provide dozens to over a hundred cores, optimized for massively parallel workloads common in AI training and scientific simulations.

  2. GPU-CPU Integration
    Grace is designed to work closely with NVIDIA GPUs via NVLink, providing extremely high bandwidth and low latency connections. This allows AI workloads to efficiently move data between CPU and GPU.

  3. Memory Bandwidth and Capacity
    Grace supports DDR5 and high-bandwidth memory configurations to handle large datasets and memory-intensive applications, a crucial feature for AI and HPC workloads.

  4. Energy Efficiency
    Despite high core counts and memory bandwidth, Grace CPUs are designed for energy efficiency in data center environments.

  5. Security and Reliability
    Includes modern hardware security features for cloud and HPC environments, ensuring protected data and reliable performance.

Use Cases

  • AI Training and Inference: Paired with NVIDIA GPUs, Grace CPUs accelerate AI model training and inference pipelines.
  • High-Performance Computing (HPC): Ideal for simulations, scientific computations, and workloads requiring high parallelism.
  • Cloud Data Centers: Designed for large-scale cloud deployments where CPU-GPU integration and energy efficiency are critical.
  • Supercomputers: Grace CPUs are integrated into supercomputers to handle massive AI workloads and large-scale scientific data.

Comparison to Other ARM CPUs

  • Versus AWS Graviton4 / Ampere Altra Max: Grace focuses on GPU-accelerated AI workloads and high memory bandwidth, while Graviton and Altra Max target general-purpose server and cloud workloads.
  • Unique GPU Connectivity: NVLink integration provides unparalleled bandwidth for CPU-GPU communication, making it uniquely suited for AI and HPC.

Summary

The NVIDIA Grace CPU represents a new class of ARM-based processors optimized for AI, HPC, and data-intensive workloads. Its combination of high core counts, GPU integration, and high memory bandwidth makes it a powerful CPU for next-generation supercomputing and AI systems.