Back to Glossary
GPU Cloud
InfiniBand
InfiniBand is a high-speed networking standard used in HPC and AI GPU clusters.
Definition
InfiniBand is a high-performance, low-latency networking standard developed by the InfiniBand Trade Association (IBTA). It is the de facto standard for GPU cluster interconnects in AI training, offering bandwidths up to 400 Gb/s (NDR) and latencies under 1 microsecond. InfiniBand supports Remote Direct Memory Access (RDMA), enabling GPU-to-GPU data transfers without CPU involvement. Harch Corp uses NVIDIA Quantum-2 InfiniBand switches for its GPU clusters, providing 400G per port for maximum distributed training throughput.
Related Keywords
infinibandinfiniband ndrrdmagpu networking400g infiniband
Related Terms
GPU Cluster
A GPU cluster is a network of GPUs working together for distributed AI training and HPC workloads.
Distributed Training
Distributed training splits AI model training across multiple GPUs or nodes for faster convergence.
RDMA
RDMA (Remote Direct Memory Access) allows direct memory access between GPUs without CPU overhead.