Phoenix enables GPU direct storage backend for LMCache

blog.lmcache.ai13 min read

Phoenix is an open-source, full-stack direct storage I/O layer designed to bypass host memory and write NVMe SSD data directly into GPU memory in a single hop. Tested as a direct storage backend for LMCache, the system significantly reduced time to first token and increased concurrent request capacity under heavy agentic workloads compared to traditional DRAM and tiered storage setups. The project, originating from Tencent's xPU-IO initiative, supports standard Linux I/O stacks and multiple hardware accelerators. Open article →