The NVIDIA GH200 Grace Hopper™ Superchips

Elevated Performance and Swifter Memory

Exceptional Bandwidth for Enhanced Compute Efficiency

The NVIDIA GH200 Grace Hopper™ Superchip represents a groundbreaking accelerated CPU meticulously crafted for massive-scale AI and HPC applications. This Superchip achieves a remarkable increase of up to 10X in performance for applications handling terabytes of data, empowering scientists and researchers to unlock unprecedented solutions for the world’s most intricate challenges.

The NVIDIA GH200 Grace Hopper Superchip integrates the NVIDIA Grace™ and Hopper™ architectures through NVIDIA® NVLink®-C2C, establishing a CPU+GPU coherent memory model designed to accelerate AI and HPC applications.

  1. Engineered for massive-scale AI and HPC, the CPU+GPU configuration is purpose-built.
  2. Features a groundbreaking coherent interface with a speed of 900 gigabytes per second (GB/s), surpassing PCIe Gen5 speeds by 7X.
  3. Elevates accelerated computing and generative AI capabilities with HBM3 and HBM3e GPU memory.
  4. Compatible with all NVIDIA software stacks and platforms, encompassing NVIDIA AI Enterprise, HPC SDK, and Omniverse™, providing seamless integration.
Harnessing power through the Grace CPU

The NVIDIA Grace CPU is meticulously crafted to deliver high single-threaded performance, exceptional memory bandwidth, and remarkable data-movement capabilities. This design achieves an optimal equilibrium between performance and energy efficiency.

Optimal performance and swiftness with GH200

The GH200 is poised to deliver a substantial increase of up to 10x in performance compared to the NVIDIA A100, particularly for applications handling terabytes of data. This advancement is set to empower scientists and researchers to attain groundbreaking solutions for the world’s most intricate problems.

The potency of unified memory architecture

The NVIDIA GH200 introduces a remarkable 7X increase in bandwidth between the CPU and GPU, a notable enhancement compared to conventional accelerated systems. This connection facilitates unified cache coherence, creating a single memory address space that amalgamates system and HBM GPU memory, simplifying programmability.

Data from NVIDIA GH200