What Is Bisection Bandwidth?_
Bisection bandwidth measures the worst-case throughput between any two partitions of a network, often used to gauge fabric balance and congestion resistance. In GPU clusters, it is a key metric for determining whether the network can sustain all-to-all communication patterns without bottlenecking. A higher bisection bandwidth relative to the aggregate GPU memory bandwidth indicates a more balanced design for distributed training.
Technical Details
Bisection bandwidth is computed by cutting the network into two equal halves and summing the link capacities crossing that cut, then taking the minimum over all such cuts. In a leaf-spine topology, the bisection bandwidth is limited by the number and speed of spine switches and their uplinks to leaf switches. For systems using NVLink (a copper-based interconnect typically confined to a single rack or node), the GPU-to-GPU bandwidth inside the rack is much higher than the scale-out fabric (fiber-based, often using MPO connectors) bisection bandwidth. The latter often becomes the bottleneck for multi-rack jobs. The exact ratio of bisection bandwidth to aggregate GPU memory bandwidth depends on the specific switch count, link speeds, and oversubscription ratio chosen per the OEM design.
How Leviathan Systems Works with Bisection Bandwidth
When commissioning a multi-rack GPU cluster, our field crew verifies that the fabric cabling and switch configurations meet the specified bisection bandwidth, often by running microbenchmarks that saturate the network. A low bisection bandwidth can cause training jobs to stall, so we check that spine-to-leaf link counts and cable types match the design intent.
Related service
Network Testing & Commissioning →