Huawei Debuts 800G AI Network Cards and KVU SSDs at WAIC 2026
At the WAIC 2026 conference, Huawei showcased its full range of computing modules, featuring the domestic-first 800G SP560 AI network card and a specialized KVU SSD solution. The KVU SSD is designed to optimize KV Cache storage offloading for AI training and inference. These hardware developments address the critical "memory wall" bottleneck in AI infrastructure by offloading memory-intensive KV Cache to high-speed storage and enabling ultra-large-scale networking of up to 100,000 nodes. This helps lower computing costs and improve NPU utilization during large language model (LLM) deployment. The 800G SP560 network card supports the Lingqu and PCIe dual-bus interface, allowing direct data transfer to GPU/NPU memory to reduce latency. Meanwhile, the KVU SSD leverages the Lingqu protocol and onboard CPU resources to achieve sequential read speeds of up to 40GB/s and random reads of 5 million IOPS.
## BACKGROUND
In large language model inference, the Key-Value (KV) Cache stores past keys and values to avoid redundant computations, but it quickly consumes massive amounts of expensive GPU memory. To bypass this "memory wall," systems offload KV Cache to high-speed SSDs. Huawei's proprietary "Lingqu" is an interconnect protocol designed for supernodes to support high-bandwidth communication between computing and storage units.