Ray 2.58.0 Posts 60% Throughput Gain

This is a meaningful infrastructure result, but it is a workload benchmark—not a universal speedup guarantee.

• –The test produced 70 million captions in 95 minutes, highlighting Ray Data’s fit for large-scale video and physical-AI data curation.
• –Gains across streaming, inference, and shuffle suggest improvements throughout the data plane, not just in one optimized stage.
• –Storage and caching were critical: CoreWeave’s LOTA cache loaded a 17 GB Qwen3-VL model in roughly two seconds.
• –The results may not generalize beyond Blackwell GPUs, CoreWeave storage, and the tested pipeline configuration.
• –Teams should benchmark their own GPU, storage, and shuffle topology before planning around these percentages.