Huawei has detailed how its UnifiedBus interconnect lets storage join the high-speed east-west traffic inside NPU-based AI clusters — a design Huawei says can address poor utilisation caused by communication bottlenecks. For UK buyers weighing storage solutions outside the Nvidia ecosystem, it's one of the most fully specified non-US alternative stacks described by Huawei.
View the data behind this chart
| Spectrum-X (NVMe link) | NVLink (Blackwell GPU) | UnifiedBus M900 link | |
|---|---|---|---|
| Bandwidth | 100GB/s | 1800GB/s | 40000GB/s |
Why GPU clusters are starving on their own compute
Huawei's briefing describes a scaling problem familiar to anyone operating large accelerator fleets: the bigger the cluster, the smaller the share of its theoretical capacity that actually gets used by models. Huawei says as little as 20% is utilised, with substantial compute idle during data communications.
Huawei attributes much of this to KV cache: cached key/value attention states generated during model processing, including interactions between large language models and AI agents, which it says now "far exceeds memory capacity" on individual accelerators and is driving demand across every storage tier in a system. A 10-trillion-parameter model requires enough intermediate data during training to exceed the memory capacity of a single accelerator under traditional architectures, forcing constant movement of data between processors, memory and storage.
What UnifiedBus actually changes
UnifiedBus is Huawei's answer: a single interconnect scheme that folds more than ten separate protocols into one, lifting interconnect bandwidth from the 100 GB/s class to the TB/s class and cutting round-trip latency from around 7 microseconds to 2 microseconds. Crucially, it also supports unified global memory addressing within SuperPoDs, rather than treating storage as a separate fabric reached through a CPU.
- •Unified protocol and memory semantics across CPUs, NPUs, memory and SSDs
- •Heterogeneous compute collaboration with Attention-FFN disaggregation for flexible CPU-NPU mixes
- •Tiered, globally pooled storage using DDR memory to cache activations — Huawei says this is intended to reduce HBM pressure
- •Optoelectronic interconnect scaling from a single cabinet to a million-NPU SuperCluster
Storage gets a seat at the GPU table
The practical consequence is that Huawei's OceanStor M900 can connect into UnifiedBus via an optical switch and act as a shared L3.5 KV cache tier, offering up to 4PB of shared L3.5 KV cache, with 40TB/s aggregate bandwidth per cluster (Huawei separately claims up to 64PB of total context-memory capacity). Huawei's own comparison describes an Nvidia configuration in which evicted KV cache stored on NVMe drives reaches GPU HBM through BlueField-mediated Spectrum-X Ethernet links running at roughly 100 GB/s — a fraction of the 1.8TB/s Huawei's comparison cites for Blackwell NVLink, or the 3TB/s cited for Vera Rubin.
By routing storage traffic through the same high-speed fabric used for processor-to-processor links, Huawei argues that NVMe-stored cache data can reach its NPUs faster than equivalent data can reach Nvidia GPUs today. It's a vendor claim rather than an independently benchmarked result — no independent comparative benchmark is provided, and aggregate cluster bandwidth is not equivalent to per-request or end-to-end latency — but it identifies a genuine architectural gap worth testing before committing to either path.
The UK angle: an alternative for buyers locked out of Nvidia supply
For UK organisations facing Nvidia allocation constraints, long lead times, or a policy preference for non-US supply chains on sovereign AI projects, UnifiedBus could interest UK organisations seeking alternatives: Huawei has now described a named stack spanning accelerators (TaiShan 950 and Atlas 960 SuperPoDs), interconnect (LinkDevice, UBG switch) and storage (OceanStor M900) working as one system — though UK availability, procurement suitability and compliance remain unverified. The LinkDevice alone claims 176 ports at 1.6 Tbit/s each for 280 Tbit/s of all-optical interconnect, while the UBG switch's 1,024 radix is pitched at supporting SuperClusters up to a million NPUs.
Huawei also says it has scaled the idea down: two Atlas 650E air-cooled servers can be switch-free linked via UnifiedBus across 16 NPUs, which Huawei says forms a small SuperPoD aimed at SMEs wanting to run trillion-parameter models on-premises rather than renting cloud GPU capacity — a relevant option for buyers using our storage solution finder to weigh on-prem versus cloud AI infrastructure.
What Huawei hasn't shown UK buyers yet
The gaps matter as much as the claims. The material reviewed provides no UK pricing, confirmed UK availability, named UK reference deployments or independent comparative benchmarks of UnifiedBus latency, sustained bandwidth or failure behaviour against NVLink, InfiniBand or Ethernet fabrics. Every figure in this analysis — the 20 percent utilisation estimate, the 40 TB/s M900 bandwidth, the 2-microsecond RTT — comes from Huawei's own briefings.
There's also a structural trade-off buyers should weigh carefully: a UnifiedBus-based deployment means accelerators, interconnect, switches, storage and software all come from one vendor. That's a different kind of lock-in risk to the one UK teams already navigate with licensing shifts like the VMware alternatives question post-Broadcom — concentrated dependency, just on a different supplier.
View the data behind this chart
| Huawei UnifiedBus | Nvidia NVLink/Spectrum-X | Traditional Ethernet | |
|---|---|---|---|
| Interconnect bandwidth | TB/s class | Up to 3 TB/s (GPU-GPU) | ~100 GB/s typical |
| Round-trip latency | ~2 microseconds | Varies by fabric | Higher, multi-hop |
| Storage access path | Direct peer access to SSD/mem | Via BlueField DPU + Ethernet | Via separate SAN fabric |
| Max cluster scale | Up to 1 million NPUs | Vendor-dependent SuperPod | Limited by fabric design |
| KV cache bandwidth | 40 TB/s aggregate (M900) | ~100 GB/s (NVMe to GPU) | Not purpose-built |
Practical steps before committing budget
Buyers evaluating this path should treat Huawei's figures as a starting hypothesis to be tested, not a procurement spec. That means seeking proof-of-concept access, demanding independent latency and throughput testing under representative KV cache loads, and modelling total cost of ownership — including how a pooled, UB-connected storage tier interacts with existing protected capacity, which is where tools like our RAID calculator and IT finance calculator earn their keep during the business case stage.
Anyone currently retiring legacy GPU-adjacent storage as part of an architecture refresh should also run it through our storage end-of-life checker before assuming a straight swap-in is possible. If you're assessing whether a non-Nvidia, non-US stack fits your sovereign AI or compliance requirements, it's worth a structured conversation — talk to a Servnet engineer before the business case is finalised.
- 01Blocks & Files — Huawei's Unified Bus lets storage in on east-west GPU cluster traffic · 18 November 2026
- 02The Register — AI inference is running into a memory bottleneck: Huawei wants to tackle it with OceanStor M900 · 19 October 2026
- 03Data Center Dynamics — Huawei unveils new UnifiedBus computing architecture for SuperPoDs and clusters · 12 November 2026
- 04Computer Weekly — Huawei expands vision for agentic world of infrastructure · 10 November 2026
- 05TechRadar Pro — Huawei Ascend 960 SuperPoD aims to surpass Nvidia AI hardware in China · 28 October 2026
- 06Blocks & Files — Huawei's OceanStor KV cache storage for hyper-scale AI data centers · 18 September 2026
- 07theregister.com
- 08datacenterdynamics.com
