# Huawei’s Unified Bus lets storage in on east-west GPU cluster traffic

> Source: <https://www.blocksandfiles.com/flash/2026/11/18/huaweis-unified-bus-lets-storage-in-on-east-west-gpu-cluster-traffic/5310818>
> Published: 2026-11-18 14:31:34+00:00

**Key takeaway:** Huawei's UnifiedBus merges more than ten interconnect protocols into one, lifting bandwidth from 100 GB/s to TB/s and cutting RTT from 7 to 2 microseconds while pooling DDR as NPU alternative memory to lower per-NPU HBM needs for 10 trillion-parameter training.

# Huawei’s Unified Bus lets storage in on east-west GPU cluster traffic

[Huawei](https://www.blocksandfiles.com/ai-ml/2026/10/19/huaweis-bridgehead-embiggening-in-the-eu/5303880) has a UnifiedBus (UB) Interconnect architecture facilitating collaboration between clusters and SuperPoDs containing its [NPUs](https://www.theregister.com/systems/2026/09/17/huaweis-next-gen-ascend-npus-could-become-chinas-best-option/5297320) (neural processing units), Huawei’s term for highly parallel GPU-like accelerators, and enabling KV cache storage access.

The company says there is an interconnect bottleneck affecting AI workload processing. Compute resource utilization declines as a cluster scales. As little as 20 percent of a 100,000-NPU cluster's computing capacity is actually utilized by models, with a vast amount of compute sitting idle during data communications. A 10 trillion-parameter model, which requires massive amounts of intermediate data to train, far exceeds the capacity of the memory on a single accelerator under traditional architectures.

AI agent and model interactions need dynamically varying amounts of CPU and NPU resources. Huawei says: “The resulting interactions between these processors also generate KV cache that far exceeds memory capacity, which is driving up demand for data storage across compute systems and tiers. All of this means interconnection and communications between components within a cluster have become a key performance bottleneck for computing systems.

Instead of having a very high-speed east-west link between processing clusters, with no linkage to slower north-south networks, Huawei has designed its UnifiedBus networking scheme to combine elements of the two.

UnifiedBus-powered computing systems have four main features:

- Unified protocol and memory semantics: More than ten interconnect protocols have been combined under the UnifiedBus protocol, which increases interconnect bandwidth from 100 GB/s level to TB/s level and reduces round-trip time (RTT) latency from 7 microseconds to 2 microseconds. This protocol also enables unified global memory addressing within SuperPoDs.
- Heterogeneous compute collaboration: UnifiedBus directly interconnects CPUs, NPUs, memory, and solid-state drives (SSDs) to enable decentralized, peer-to-peer access between them and flexible CPU-NPU mixing. In addition, tiered hardware acceleration for Transformer enables Attention-FFN disaggregation (AFD).
- Tiered storage with global pooling: UnifiedBus uses hybrid-media resource pooling to support the caching of activations. Double Data Rate (DDR) memory serves as the alternative memory for NPUs, reducing latency for search, recommendation, and advertising services, and doubling the performance of vector retrieval for 100 billion-level data items with thousands of dimensions. This also reduces the per-NPU HBM capacity requirements for the training of 10 trillion-parameter models, and improves cluster model FLOPs utilization (MFU).
- Optoelectronic interconnect and flexible networking: UnifiedBus serves as a global ultra-high bandwidth, ultra-low latency "data highway", enabling flexible scaling of compute

The UnifiedBus architecture covers interconnection within cabinets, between cabinets, and across clusters, enabling elastic scaling from a single cabinet to a million-NPU cluster.

- Within cabinets, the UnifiedBus LinkBlade has a cable-free design, which reduces copper cabling within a 4,096-NPU SuperPoD by about 196 kilometers.
- Between cabinets, the UnifiedBus LinkDevice supports 176 ports, with 1.6 Tbit/s bandwidth per port, delivering 280 Tbit/s all-optical interconnection. This is the industry's highest-bandwidth, highest-port-count, high-speed bus-protocol interconnect device, with an RTT latency as low as 2 microseconds.
- Across clusters, the UnifiedBus photonic UBG switch provides a massive radix fan-out capacity of up to 1,024, meaning it can support a million-NPU SuperCluster and evolution towards supporting tens of trillions of model parameters.

Huawei has a UnifiedBus-powered agentic AI SuperCluster, which is comprised of TaiShan 950 and Atlas 960 SuperPoDs, OceanStor M900 memory storage, and a Xinghe UBG switch.

It has also applied the UnifiedBus technology to its compute appliances. Two Atlas 650E air-cooled servers can be directly linked through UnifiedBus to enable switch-free interconnect across 16 NPUs, so they work as a single unit. They can scale to form a small-scale SuperPoD that enables small and medium-sized enterprises to run trillion-parameter models locally.

##### **Cross-cluster KV cache sharing**

As we understand it, in the [Nvidia GPU world](https://www.blocksandfiles.com/2026/01/12/nvidias-basic-context-memory-extension-infrastructure/4090541), HBM-evicted KV cache contents needed to be retained are pushed down through progressively slower memory and storage tiers to NVMe SSDs. Reloading this data from them for large models is faster than recomputing the KV cache values but still slow. Models can be transferred between GPUs with NVLink, Nvidia’s east-west network operating at 1.8 TB/s with Blackwell GPUs and 3 TB/s with Vera Rubin ones. But NVMe SSDs in Nvidia’s 3.5 tier send data to GPU’s HBM using BlueField-4 mediated Spectrum-X Ethernet links operating at ~100 GB/s, much, much slower.

Wu (Jeff) Junjie, Huawei’s Data Storage Product Line CMO, said in a briefing that Huawei’s [M900](https://www.blocksandfiles.com/flash/2026/09/18/huaweis-oceanstor-kv-cache-storage-for-hyper-scale-ai-data-centers/5297405) context memory storage product can link to the UB via an optical switch. The M900 provides up to 4 PB of shared L3.5 KV cache capacity and 40 TB/sec of aggregate bandwidth per cluster. 

This is faster than a Spectrum-X link and means, as we understand it, that NVMe drive-stored KV cache contents can be sent to Huawei’s NPUs faster than they can be sent to Nvidia’s GPUs.

##### Comment

We daresay suggest that, when Nvidia was first thinking about GPU-to-GPU linkage, and seeing a need for very high speed connectivity, the evicted KV cache data transmission load was not seen as much of a factor to consider. Using BlueField DPUs to offload the GPU server’s CPU and manage this storage traffic over Ethernet was then quite sufficient.

Step forward to modern day times and, seen from Huawei’s point of view, it’s no longer the case. Evicted KV cache data needs to be shared between its NPUs at high speed to keep their utilization high and opening up UB connectivity to storage like the M900 achieves that.

How might Nvidia respond is an interesting question.

##### **Bootnote**

Attention is a mechanism that captures contextual relationships, while FFNs (Feed-Forward Networks) handle layer-by-layer feature transformation. They are both core computational modules of the Transformer architecture.

Attention-FFN disaggregation (AFD) decouples Attention and FFN and deploys them onto different compute units to enable flexible CPU-NPU mixes for different workloads. This improves inference efficiency and resource utilization. AFD is particularly valuable for agentic AI which requires frequent CPU-NPU collaboration.
