Huawei is pitching UnifiedBus as a way to connect processors, memory and storage more tightly inside large AI clusters, with the aim of reducing bottlenecks as models and their working data grow. The practical prize is faster access to cached information that would otherwise leave accelerators waiting.
Huawei AI Watch analysis
What happened
Blocks & Files reports that Huawei’s UnifiedBus combines more than ten interconnect protocols and provides unified memory addressing across its SuperPoD systems. Huawei says the design can raise interconnect bandwidth from the 100 GB/s range to the terabyte-per-second range, while cutting round-trip latency from seven microseconds to two.
The architecture connects CPUs, NPUs, memory and SSDs, and supports a shared storage tier for KV cache, the intermediate data used during model inference. Huawei says its OceanStor M900 can provide up to 4 PB of shared KV-cache capacity and 40 TB/s of aggregate bandwidth per cluster. Those are company figures reported by Blocks & Files, not independent test results. The report also describes a system built from Huawei’s TaiShan 950 and Atlas 960 SuperPoDs, M900 storage and a Xinghe optical switch. Read the Blocks & Files report.
Why it matters
Large models can generate more working data than a single accelerator can hold. Moving that data between compute and storage can become a serious bottleneck, particularly when AI agents repeatedly shift work between processors. Huawei’s pitch is that storage should be part of the high-speed cluster fabric, rather than waiting at the end of a slower connection.
That makes this more than a new cable story. If the claimed bandwidth and latency translate into real workloads, the design could affect how efficiently operators serve large models and how much accelerator capacity sits idle while data moves.
Our read
The interesting idea is the closer coupling of compute and storage, especially for the KV cache. But Huawei’s headline specifications are not the same thing as independent evidence of faster model performance. The useful next step is clear: publish workload-level results that show whether the architecture improves throughput, latency or utilisation in practice.
What to watch
- Independent measurements of UnifiedBus performance on real training and inference workloads.
- How Huawei’s shared KV-cache storage performs as cluster size and concurrent demand increase.
- Whether operators can deploy the architecture beyond Huawei’s own integrated systems.
Discussion spark: For large AI systems, should operators prioritise faster links between accelerators, or put more effort into keeping working data close to where it is used?
Sources and evidence
- Huawei’s Unified Bus lets storage in on east-west GPU cluster traffic – Blocks & Files (18 November 2026, 14:31 UTC)
Independent WittyWires tracker for public updates about Huawei AI. Not affiliated with or endorsed by Huawei AI; this is not an official account.