S3 to GPU at 60Gbps: Why Parquet is Bottlenecking Your Data

Parquet and traditional NVMe data pipelines are slowing down your GPU analytics and ML training workloads. In this deep dive, Onur Satici (SpiralDB / Linux Foundation) breaks down how Vortex - an open-source, fast-path columnar file format - streams data directly from S3 to GPUs at up to 60 Gbps, eliminating CPU decompression bottlenecks and saving GPU compute cycles. Discover how Vortex decouples logical and physical types, leverages cascading encodings for compute on compressed data, and enables zero-copy execution with Apache Arrow, DataFusion, DuckDB, and Polars. 📌 Timestamps: 00:00 - Live Demo: Visualizing a 4K 60fps Scan from S3 to GPU 02:08 - The Problem: GPU Movement Tax vs. Decision Tax 04:09 - Introducing Vortex: Parquet vs. Vortex Architecture 05:27 - Vortex File Spec: Postscript, Footers, and Segment Alignment 06:55 - How Layouts and Zone-Maps Eliminate Unnecessary Disk I/O 09:20 - Cascading Encodings: DictArray, RunEnd, and Bitpacking 11:58 - Executing Compute directly on Compressed Data (Zero-Copy) 12:53 - Scan Architecture: Byte Coalescing for CPU and GPU 14:41 - Bypassing CPU Memcopies: Buffer Pools, kTLS, and RDMA 17:42 - Advanced Demo: Custom GPU Expressions & Zone Pruning 19:50 - Edge Cases: Multi-GPU RDMA & Distributed Aggregations 21:30 - Q&A: Video Codecs, Database Engine Support (DuckDB, Iceberg), and Benchmarks 🔗 Transcript & slides available on InfoQ: https://bit.ly/4hfh0Ah #SoftwareArchitecture #DataEngineering #GPUAcceleration #Parquet #Vortex