NVLink

NVIDIA's proprietary high-speed GPU-to-GPU interconnect technology that enables direct communication between GPUs at much higher bandwidth than PCIe. NVLink provides up to 600 GB/s bidirectional bandwidth (on H100), compared to PCIe Gen4's approximately 64 GB/s. Essential for multi-GPU training and large models that need to share data across GPUs frequently. In cloud instances, NVLink topology matters: AWS p4d instances connect 8x A100 GPUs via NVLink in a hybrid cube-mesh topology, while p5 instances use NVLink for H100s. NVLink-enabled instances typically cost more but deliver significantly better performance for distributed training workloads.