NVLink

Connectivity

DEFINITION

NVIDIA's high-speed interconnect technology for direct GPU-to-GPU communication, bypassing the CPU and PCIe bus.

OVERVIEW

NVLink is a high-bandwidth, low-latency interconnect that allows multiple GPUs to communicate directly with each other at speeds much faster than PCIe, enabling efficient multi-GPU training and inference.

TECHNICAL DETAILS

NVLink provides bidirectional bandwidth of 600 GB/s (NVLink 3.0) or 900 GB/s (NVLink 4.0), compared to PCIe 4.0's ~64 GB/s. This allows GPUs to share memory and data much more efficiently. NVLink is essential for model parallelism and large-scale distributed training where GPUs need to synchronize gradients frequently.

COMMON USE CASES

  • Multi-GPU training of large models
  • Model parallelism across multiple GPUs
  • High-performance computing clusters
  • Real-time multi-GPU inference
  • Distributed deep learning