Blog

MLPerf Storage v3.0: Alluxio's Distributed Cache Delivers 147 GiB/s Llama 3 Checkpoint Writes

Alluxio’s MLPerf Storage v3.0 results demonstrate that organizations can keep persistent AI data in Amazon S3 while delivering high-performance training and checkpointing through a distributed cache close to compute. Alluxio achieved 147.03 GiB/s checkpoint write bandwidth for Llama 3 405B and near-linear scaling from one to 32 workers.

‍

Diagnose & Fix Slow Distributed Training

Got periodic drops in GPU utilization? GPU Stalls? Training capacity grinding to a halt? Learn how checkpoint writes could be the cause of your suddent, yet periodic drops in training performance.

20x Faster Training Data Reads with Alluxio and Ray on Anyscale: A Cross-Region Benchmark

Alluxio and Anyscale benchmark achieves 20x faster cross-region data reads for AI training workloads on GCS.

Sign-up for a Live Demo or Book a Meeting with a Solutions Engineer

Request a demo