uGDS is the first production-oriented user-space GPU Direct Storage library. The CPU constructs NVMe commands and the SSD DMAs data directly to/from GPU memory over PCIe, completely bypassing the kernel NVMe driver.
With a GDS-compatible API, uGDS achieves up to 5.3x higher bandwidth and 108x lower 4KB latency (5.2μs vs 561μs) compared to NVIDIA GDS. It supports both NVIDIA CUDA and AMD ROCm backends, and has been integrated as the storage backend for LMCache, delivering 2x TTFT reduction and 2x throughput improvement in vLLM end-to-end benchmarks.