InfiniBand
Networking & InterconnectA high-performance networking fabric that connects GPU nodes at low latency and high bandwidth, the interconnect standard for serious AI training clusters.
Eleveight AI's cluster uses the NVIDIA InfiniBand XDR networking platform, as part of its NVIDIA Reference Architecture compliance, keeping inter-node GPU communication fast enough for large distributed training runs.
Overview
InfiniBand is a networking technology born in high-performance computing and now the dominant interconnect for large AI training clusters. Its appeal is a combination ordinary networks struggle to match: latency measured in microseconds alongside very high bandwidth between nodes. Where standard Ethernet was designed for general traffic, InfiniBand was built from the start for tightly-coupled machines that must move data constantly.
How it works
It uses a switched-fabric topology, with GPU servers connecting into InfiniBand switches that route traffic between them. Its defining feature is Remote Direct Memory Access, or RDMA, which lets one node read from or write to another node's memory directly, bypassing the CPU and operating system entirely. That shortcut strips out overhead and is much of why its latency is so low. The XDR generation used in Eleveight AI's cluster delivers the highest InfiniBand bandwidth available.
Why it matters
During training spread across dozens or hundreds of nodes, the GPUs must continually exchange gradient information to stay synchronized, and the network carrying that exchange can easily become the limiting factor. InfiniBand's low latency and high bandwidth keep the accelerators fed and the cluster efficient, so adding more nodes actually translates into faster training rather than wasted capacity.
Use cases
- Connecting GPU nodes in large-scale AI training clusters
- Enabling efficient gradient synchronization
- High-throughput storage access for training data pipelines
- AI factory networking backbone