HN Beats MLPerf with Modern Storage for KV Offload and LLM Training
gpu training
| Source: HN | Original article
An open‑source storage system using Rust compio’s thread‑per‑core design and on‑GPU compression claims to outpace MLPerf benchmarks for KV offload and LLM training.
OpenLake, a storage system built for AI workloads, announced that it has topped the MLPerf Storage v3.0 benchmark for large‑language‑model (LLM) training. In the 8‑billion‑parameter checkpoint write test, OpenLake recorded the highest read and write bandwidth among the five S3‑compatible submissions, delivering 11.55 GiB/s of reads and 6.72 GiB/s of writes. The result was achieved with a Rust‑based “compio” engine that follows a thread‑per‑core model and performs on‑GPU compression, while leveraging Linux’s io_uring and NVIDIA’s GPUDirect Storage to keep latency low.
The achievement matters because storage has become a critical bottleneck in modern LLM pipelines, especially when key‑value (KV) caches are offloaded from GPU memory to persistent media. Faster, lower‑latency I/O can shrink training cycles, reduce hardware costs and enable larger models without proportionally larger GPU memory pools. OpenLake’s performance demonstrates that software‑defined storage, when tightly coupled to the GPU, can rival or surpass traditional cloud object stores in the most demanding AI scenarios.
What to watch next is whether cloud providers and enterprise AI teams adopt OpenLake’s approach, and how the system fares in upcoming MLPerf rounds that will test broader workloads such as inference and multi‑node training. The broader AI community will also be looking for open‑source or commercial integrations that expose OpenLake’s io_uring/GPUDirect stack to popular frameworks. As we noted earlier in “Architecting memory and storage in the AI era” (2026‑09‑05), advances in storage architecture are now as pivotal as GPU or model innovations for scaling the next generation of AI.
Sources
Back to AIPULSEN