NVIDIA Grace CPU

Solving the largest AI and HPC problems requires high-capacity and high-bandwidth memory (HBM). The fourth-generation NVIDIA NVLink-C2C delivers 900 gigabytes per second (GB/s) of bidirectional bandwidth between the NVIDIA Grace CPU and NVIDIA GPUs. The connection provides a unified, cache-coherent memory address space that combines system and HBM GPU memory for simplified programmability. This coherent, high-bandwidth connection between CPU and GPUs is key to accelerating tomorrow’s most complex problems.

NVIDIA Grace is the first server CPU to harness LPDDR5X memory with server-class reliability through mechanisms like error-correcting code (ECC) to meet the demands of the data center, while delivering 2X the memory bandwidth and up to 10X better energy efficiency compared to today’s server memory. The LPDDR5X solution coupled with NVIDIA Grace’s large high-performance last-level cache delivers the bandwidth necessary for large models while reducing system power to maximize performance for next-generation workloads. 

As the parallel compute capabilities of GPUs continue to advance, workloads can still be gated by serial tasks run on the CPU. A fast and efficient CPU is a critical component of system design to enable maximum workload acceleration. The NVIDIA Grace CPU integrates Arm Neoverse V2 cores with the NVIDIA Scalable Coherency Fabric to deliver high performance in a power-efficient design, making it easier for scientists and researchers to do their life’s work.

Generative AI is memory and compute intensive. The NVIDIA GB200 Superchip uses 380GB of HBM memory, delivering over 4.5X the GPU memory bandwidth of the NVIDIA H100 Tensor Core GPU. The high-bandwidth memory in Grace Blackwell is connected to CPU memory over NVLink-C2C to provide almost 860GB of fast-access memory to the GPU, delivering the memory capacity and bandwidth needed to handle the world’s most complex generative AI and accelerated computing workloads.