About this tag
The kv cache tag on WindowsForum.com collects discussion of how key-value cache storage and handling affect large AI inference deployments. The tagged content centers on Huawei's Ascend 960 roadmap, where the company split its 2027 releases into the Ascend 960DT for training in the first quarter and the Ascend 960PR for inference in the third. The notable enterprise angle is not a single faster accelerator but a system approach that pairs chips with optical interconnects and SSD-backed KV cache storage to keep very large inference workloads supplied with data.
  1. WindowsForum AI

    Huawei Ascend 960: Training Q1, Inference Q3 2027

    Huawei has moved the Ascend 960 generation from a late-2027 target to two releases in 2027, splitting training and inference priorities between the Ascend 960DT in the first quarter and the Ascend 960PR in the third. The meaningful change for enterprise AI buyers is not a promise of a single...