Pandaily reported on 23 October that Huawei presented lab results for its Stellar AI Fabric at the Advanced Storage Conference in Shanghai on 22 October. Huawei said moving a model’s KV Cache to storage increased inference throughput in tests using GLM-5.2. The Open Data Center Committee organised the conference.
The KV Cache holds intermediate results so a model can reuse work on earlier tokens. Huawei says long, repeated interactions make the cache grow and increase traffic between machines. Its Hyper-Converged Fabric architecture coordinates storage and network traffic, including scheduling tailored to cache transfers. Huawei says the network layer alone improves storage input-output performance by more than 30%.
Continuing a long conversation could create more cached results to move between machines, while reusing those results would spare the model from recalculating earlier tokens.
In Huawei’s lab tests, direct cache offloading delivered 7.3 times the inference throughput of a conventional memory-based approach. The company said coordinated storage and network scheduling improved inference efficiency by a further 20%. Huawei introduced the architecture in September, when AI Affairs reported on its network-interactive data-centre design. Pandaily reported that Huawei did not publish the full test configuration.
.jpg){kind=link}