Presentation: From S3 to GPU in One Copy: Rethinking Data Loading for ML Training
60 Gbps from S3 to GPU, zero-copy: Vortex eliminates the data-loading bottleneck killing training throughput.
Why it matters
Data loading is a hidden tax on ML training economics. Vortex, a Linux Foundation columnar format, claims to stream cloud object storage directly to GPUs without CPU/NVMe friction — a meaningful shift in how practitioners architect training pipelines and reason about compute utilization.
The key facts
12 to knowVortex: open-source columnar file format under Linux Foundation
60 Gbps streaming throughput from S3 to GPU
Zero-copy memory pipeline architecture
Cascading lightweight encodings + layout-based segment pruning
Eliminates upfront data reprocessing requirement
Targets CPU/NVMe bottlenecks in high-throughput training
Achieves up to 60 Gbps streaming from S3 to GPU
Uses cascading lightweight encodings and layout-based segment pruning
Eliminates CPU/NVMe bottlenecks
Removes need for upfront data reprocessing
Speaker: Onur Satici
Published: Sep 4, 2026 (InfoQ presentation)
Go to the source
InfoQ AI/MLinfoq.com
Publisher excerpt: Onur Satici explains how Vortex, an open-source columnar file format under the Linux Foundation, revolutionizes high-throughput data loading. He details how cascading lightweight encodings, layout-based segment pruning, and zero-copy memory pipelines eliminate CPU/NVMe bottlenecks to stream S3 data…