Parquet is bottlenecking GPU pipelines: Vortex claims 30x faster scans

The pipeline problem. The main claim is that GPUs sit idle because of how data reaches them, not because of model training. The Vortex maintainer shows a typical PyTorch ingest path: S3 to NVMe, compressed bytes decompressed on CPU, then sent over PCIe to the GPU. His argument is that CPU decompression and disk staging, not network bandwidth, are the bottleneck, and that every worker duplicates shards because Python multiprocessing avoids shared state.

The format bet. Vortex is an open-source columnar format under the Linux Foundation, positioned as an alternative to Parquet for GPU-native analytics. It separates logical types from physical encodings, uses lightweight cascading encodings rather than block compression, and stores zone-map statistics so a scan can prune segments before decompressing. The speaker's key design choice is representing data as arrays that can be queried while compressed.

The speed claims. The numbers are impressive but vendor-generated. Vortex scans from S3 to GPU at about 60 Gbps, roughly 30x faster than Parquet, with 100x faster random access. CPU scans are about 20x faster, writes about 5x faster, and compressed file size stays similar. The live demo reads 4K video as three columns, prunes a blue channel, and filters frames by a "has-bike" boolean. It is a visualization, not a video codec.

The real engineering. A large part of the talk is about avoiding memory copies. The speaker describes pinned buffer pools, KTLS kernel-side TLS decryption, and a custom HTTP client that deframes multipart requests directly into buffers ready for CUDA async copy. This is the most practical section: independent of Vortex, it explains why S3-to-GPU performance depends on copy count and buffer alignment, and how tuning these changes helps any ingest pipeline.

The adoption gap. Vortex has integrations with DataFusion, DuckDB, and Polars, and supports Arrow zero-copy through canonical encodings. Iceberg support is in progress, not shipped. The speaker says file-spec V1 is stable, but the supporting tools are young, and interoperability was the first audience concern. A Parquet-to-Vortex conversion is easy, he says, but long-term lock-in and tooling maturity are open questions.

Bottom line. Teams should not move production data to Vortex based on this talk. The bottleneck diagnosis is credible; the speed numbers need independent validation on a real workload. For GPU-heavy analytics or repeated retraining, running a small benchmark is cheap and worth doing. For everyone else, watching is optional, and treating "the format is the bottleneck" as a company-wide conclusion is probably premature.