NVIDIA Says Spectrum-X Multiplane Reroutes AI Traffic Around Failed Planes in Under 3 ms

NVIDIA is positioning Spectrum-X Multiplane as an answer to a stubborn problem in giant AI clusters: one congested or failed network path can hold up a synchronized GPU job. The company says its hardware-based plane load balancer can remove a failed network plane and redirect traffic over healthy planes in under three milliseconds.

When synchronized traffic meets static paths

The issue is less about Ethernet’s general usefulness than the traffic pattern created by distributed training. GPU collectives such as All-Reduce, All-Gather, and All-to-All produce a relatively small number of large flows that must finish together. A delayed flow becomes a straggler, leaving other GPUs waiting for the collective to complete.

Conventional Equal-Cost Multi-Path, or ECMP, routing spreads flows through static hashing. NVIDIA says that can put several large AI flows on one link while parallel links remain underused, because the hash does not respond to current congestion. It also says RoCEv2 loss-management methods can spread pauses or create head-of-line blocking during congestion.

Three controls divide the job

Spectrum-X combines three hardware controls. First, switches use per-packet adaptive routing: rather than commit a flow to a hash-selected path, they choose among paths using egress-queue conditions. NVIDIA says the switch samples queues at sub-microsecond intervals and can steer packets within hundreds of nanoseconds to reduce local hot spots.

Second, targeted congestion control is intended for endpoint incast, when several senders overload one receiver. The switch applies ECN marks only after adaptive routing can no longer absorb a growing queue; senders then use RTT probes and those marks to lower transmission rates. Third, the SuperNIC’s Plane Load Balancer selects among independent network planes using local queue feedback and per-plane congestion telemetry.

Making several fabrics appear as one

The multiplane design divides a host’s network bandwidth across physically independent fabrics. NVIDIA gives an example of an 800 Gbps host connection split into four 200 Gbps planes, each built as a shallow two-tier fat tree. The operating system and collective-communication libraries see one unified RoCE device, while the SuperNIC handles distribution and failover underneath.

NVIDIA identifies a weakness in oblivious packet spraying, which sends equal traffic to every plane without accounting for a degraded link. Its Plane Load Balancer maintains congestion-control state for each destination and plane, filters planes with congestion or failure from the eligible set, then picks the remaining plane with the shallowest local egress queue.

The performance case is tied to a failure scenario

NVIDIA’s failure example assumes a 20% switch-to-switch connectivity loss on one plane of an eight-plane network. It says traditional multiplane Ethernet falls to 80% capacity because all planes are limited by the impaired one. Spectrum-X Multiplane instead keeps seven healthy planes at full capacity and the affected plane at 80%, producing 1.2 times higher All-to-All collective bandwidth in that scenario.

NVIDIA says a two-tier multiplane topology can exceed 128,000 endpoints, while a three-tier version can reach up to 16 million. Its central proposition is narrower: plane-aware hardware balancing can preserve available capacity when one plane degrades, rather than sending equal traffic into the bottleneck.