Sugon Unveils AI Inference Storage FN Neo with Pooled KV Cache
Sugon (Zhongke Shuguang) released its updated AI inference-native storage product FN Neo at the 2026 CCF National Conference on Information Storage Technology on September 21. Based on a pooled shared architecture, the system supports cluster-level KV Cache resource scheduling, addressing the problem of isolated local caches that prevent reuse across sessions and nodes in large model and agent scenarios.
In conventional inference deployments, KV Cache is treated as private compute-node resources, which can lead to idle cache data and underutilized compute power. FN Neo turns these private caches into shared resources that can be scheduled across the inference cluster, allowing flexible reading and reuse among different model sessions and compute nodes.
The vendor said real-world testing shows a single array can deliver 160GB/s bandwidth with sub-millisecond latency, roughly twice the throughput and concurrent request performance of local NVMe solutions. The system reduces first-token response time by 73% and increases overall token throughput by 150%. It is built with 99.999% reliability, using mechanisms such as slow-disk automatic isolation and fault-free reconstruction to support 24/7 inference operations.
FN Neo offers a unified SAN/NAS/KV protocol ecosystem and is compatible with mainstream inference frameworks including vLLM, LMCache and Mooncake. Sugon said the pooled storage approach reduces duplicate storage overhead across nodes and improves utilization of expensive compute hardware such as GPUs.
