Cables hold back AI clusters more than chips: Huawei's UnifiedBus promises 3,5 times lower latency

Large clusters are constrained not by chips, but by communication between them. In Huawei's measurement of 17.09, a cluster of 100,000 NPU uses around a fifth of its capacity for compute, while the rest sits idle waiting for data.

Before and now. Cluster nodes used around ten different interconnect protocols, and every junction cost time. Huawei unified more than ten protocols into one UnifiedBus. Bandwidth rises from the 100 GB/s range to the TB/s range, while response time falls from 7 to 2 us — data waits are 3.5 times shorter.

The company does not keep the UnifiedBus software stack closed. The five services of the UB Service Core project are in the openEuler community under the MulanPSL2 license, with the code and working group on AtomGit. The architecture 2.0 reference design can be downloaded there as a standard PDF, with no registration required. Huawei claims a 30–50% application uplift on this stack.

An outside measurement. Engineer Bojie Li built his own UB transport implementation from scratch and released a preprint on 27.05.2026. His OpenURMA responds to a remote load in 500 ns, versus 2186 ns for RoCEv2 — 4.37 times faster — and has 2.80 times higher throughput. The vendor did not calculate these figures.

The first UnifiedBus system, Atlas 960E, combines up to 4096 Ascend 960 NPU. Instead of 48,000 standard 800G modules, it uses 5500 Hi-ONE optical engines, and Huawei puts the savings from that swap at more than 550 kW.

Huawei promises Atlas 960E with UnifiedBus by Q3 2027.