NVIDIA GH200 Grace Hopper Superchip

NVIDIA Grace CPU with Hopper GPU

72-core ARM CPU with H100 96GB GPU. Single module with NVLink-C2C interconnect that delivers 900 GB/s of bidirectional bandwidth between the Grace CPU and Hopper GPU.

Higher Performance and Faster Memory

Higher bandwidth system and GPU memory. 7x faster CPU to GPU transfers over NVLink.

Coherent Memory Architecture

480 GB system memory + 96 GB GPU memory. Each GH200 GPU can address up to 576 GB of memory.

The fastest network for distributed training of LLMs, foundation models & generative AI

Train large foundation models and LLMs with the fastest networking available in any cloud. Our NVIDIA Quantum-2 InfiniBand networking provides 3200 Gbps of bandwidth for each HGX H100 or H200 node.

This design is purpose built for NVIDIA GPUDirect RDMA with maximum inter-node bandwidth and minimum latency across the entire cluster.

The Lambda compute network uses a non-blocking multi-layer topology with zero oversubscription. This provides full networking bandwidth to every NVIDIA GPU in the cluster simultaneously, the optimal design for full-cluster distributed training.

Skip the CPU and take advantage of GPUDirect RDMA for the fastest distributed training

A direct communication path between NVIDIA GPUs across all nodes in your cluster using NVIDIA Quantum-2 InfiniBand.

GPUDirect RDMA provides a significant decrease in GPU-GPU communication latency and completely offloads the CPU, removing it from all GPU-GPU communications across the network.