Abstract
ML training pipelines treat data as static. Teams spend weeks preprocessing datasets into WebDataset or TFRecords, and when they want to experiment with curriculum learning or data mixing, they reprocess everything from scratch. Meanwhile, GPUs sit idle waiting for data that takes the scenic route: S3 to disk, disk to CPU, CPU decompression, and finally a copy to the GPU.
What if you could query your training data on the fly and stream it directly from object storage to GPU memory, saturating the host-to-device bandwidth?
This talk introduces Vortex, an open source columnar format designed for this world. I’ll cover how Vortex differs from Parquet - extensible encodings, GPU-native decompression, and a layout optimized for selective reads - then dive deep into a new data path: coalesced byte-range requests from S3 into pinned buffers, with H2D transfers that saturate GPU bandwidth, all in a single copy.
You’ll learn why the CPU is the bottleneck in ML data pipelines, how to remove it, and where this is heading: GPU-direct reads over RDMA that eliminate the last copy entirely.
For ML infrastructure engineers tired of the preprocessing treadmill.
Speaker
Onur Satici
Staff Engineer @SpiralDB
Onur is a Staff Engineer at SpiralDB and a core maintainer of Vortex, an open source columnar file format now part of the Linux Foundation (LF AI & Data). He focuses on high-performance data systems, GPU acceleration, and making analytical workloads faster at every layer of the stack.
Date
Tuesday Mar 17 / 01:35PM GMT ( 50 minutes )
Location
Fleming (3rd Fl.)
