OpenLake nearly doubles NVIDIA AIStore's checkpoint write speed

OpenLake reported 6.72 GiB/s of checkpoint write bandwidth on MLPerf Storage v3.0, per OpenLake's benchmark results. That nearly doubles NVIDIA AIStore's best six-node result of 3.40 GiB/s on the same Llama 3.1 8B checkpointing test.

For anyone choosing storage for GPU training clusters, checkpoint write speed sets how often a team can save model state without stalling GPUs. OpenLake's result turns that comparison into a benchmark number instead of a vendor claim.

The comparison covers five identically configured submissions in the same MLPerf round, so it isolates the storage layer from differences in training code. Both AIStore and OpenLake submitted to that round, making this a head-to-head result rather than a self-reported one. On reads, the two finished close, 11.55 GiB/s for OpenLake against 11.08 for AIStore.

Results went up September 1 and reached Show HN on September 5, as other MLPerf Storage v3.0 submissions get compared vendor to vendor.