NVSwitch is a high-speed switch chip developed by Nvidia that builds on the advanced communication capability of NVLink. It is designed to provide higher bandwidth and reduced latency for compute-intensive workloads that require multi-GPU and multi-node systems.
NVSwitch enables all-to-all GPU communication at full NVLink speed within a single node and between nodes.
NVSwitch has the following key features and capabilities:
- 18 ports of NVLink per switch, each providing 50 GB/s of bidirectional bandwidth
- Fully connected crossbar topology that allows any port to communicate with any other port at full speed
- In-network computing engines, such as NVIDIA Scalable Hierarchical Aggregation and Reduction Protocol (SHARP)™, for collective operations and multicast acceleration
- Error detection and correction mechanisms for reliable data transfer
- Low power consumption and high thermal efficiency
NVSwitch can be used for various purposes, such as:
- Accelerating AI and high-performance computing (HPC) workloads by enabling faster and larger model training, inference, and data processing.
- Enhancing GPU memory capacity and bandwidth by enabling GPU memory pooling and coherent access.
- Improving system performance and efficiency by enabling in-network computing engines, such as NVIDIA Scalable Hierarchical Aggregation and Reduction Protocol (SHARP)™, for collective operations and multicast acceleration.
NVSwitch can be combined with NVLink to create a seamless, high-bandwidth, multi-node GPU cluster that can scale up to 256 GPUs and deliver up to 57.6 TB/s of all-to-all bandwidth.
NVSwitch is also compatible with various Nvidia products, such as DGX H100, Hopper, Grace CPU, A100 GPU, and more.
