A user-space rewrite of GPU Direct Storage hit 12.12x its batch-read bandwidth and cut write latency 84.1 percent
AspireAI community · arXiv preprint
NVIDIA's GPU Direct Storage lets an SSD stream data straight into GPU memory over PCIe, which is how data-intensive AI pipelines extend limited GPU memory. Its control path, however, still runs through the kernel on every request, it enforces strict stream ordering even when a workload only needs batch ordering, and its user-space engine is closed source - which makes it hard to diagnose or schedule around.
uGDS moves the whole NVMe command lifecycle into user space, keeps the kernel only for one-time privileged setup, partitions hardware queues between synchronous and batch traffic, and exposes cuFile-style synchronous, batched and stream-ordered APIs on both NVIDIA CUDA and AMD HIP.
On an A100 with a Samsung 990 PRO, the measured numbers are large: up to 12.12x GDS's batch-read bandwidth and 30.69x its batch-write bandwidth at 8- and 4-KiB requests with batch size 128, and 84.1% lower 4-KiB synchronous write latency with two issuing threads. Its single-stream path sustains 5.49 GiB/s where GDS retains only 40.8% of its synchronous bandwidth, and plugged into LMCache it cuts median time-to-first-token by up to 62.7% while raising serving goodput 2.63x. The code is open source. Caveat: one A100 plus one consumer SSD is the whole testbed, and the paper is a technical report.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.