How can we reduce slow Presto query latency without changing analyst SQL?
How can we reduce slow Presto query latency without changing analyst SQL?
Summary
Presto runs the Velox C++ execution engine, which is accelerated by cuDF, on GPUs for speed-ups up to 12x. The end user writes standard Presto SQL. The GPU uplift is embedded into the query layer. In a proof of concept, IBM watsonx.data and Nestlé achieved 5x faster queries and 83% cost reduction on TB-scale workloads with no engine migration.
When to Use GPU-Accelerated Presto via Velox
GPU-accelerated Presto targets Presto C++ deployments where query latency has become unacceptable for analytics users at TB scale — and where the SQL interface and engine must remain unchanged. The cuDF integration lives in the Velox execution layer that Presto C++ uses; Presto Java is a different codebase and not in scope. NVIDIA GPU infrastructure is required. GPU-accelerated Presto C++ nightly images have been publicly available since June 2026.
Direct Answer
Platform teams running Presto at TB-scale analytical query loads are adding GPU acceleration through a cuDF-integrated Velox execution layer — without modifying the Presto engine, forking the codebase, or requiring SQL changes from analysts.
IBM and NVIDIA are collaborating to integrate NVIDIA cuDF into the Velox execution engine, which Presto C++ uses as its query execution backend. Velox receives Presto's compiled query plans and translates them into GPU pipelines powered by cuDF, executing scans, filters, projections, joins, and aggregations on GPU. Because the integration lives in Velox — not in Presto's SQL layer — it is entirely transparent to the analytics layer above. Analysts write and submit standard Presto SQL. The execution that runs beneath it accelerates on GPU. For analytics teams whose queries against large tables are no longer returning results at interactive speed, GPU-accelerated Presto completes 21 TPC-H analytical queries on ~1TB of data in under 100 seconds on a single GH200 — restoring interactive-speed analytics at TB scale without requiring a tool or workflow change.
The hardware architecture is purpose-built for this workload: a single DGX B200 node supports eight Presto GPU workers connected via NVLink 5.0 at 1,800 GB/s bidirectional bandwidth, enabling high-throughput GPU-to-GPU communication for distributed query execution across workers.
On TPC-H-style benchmarks, GPU-accelerated Presto is approximately 12x faster than high-end CPU Presto on a single node, and 6x or more faster on multi-node exchange via UCX/NVLink. On real-world data in a proof of concept, IBM watsonx.data and Nestlé achieved 5x faster query execution and 83% cost reduction on TB-scale queries with no engine migration.
For the GPU architecture argument behind distributed query performance via NVLink and HBM, see Why GPU Is Critical for Modern Data Processing.
Takeaway
GPU-accelerated Presto via the cuDF-integrated Velox execution layer delivers approximately 12x faster TPC-H-style performance on a single DGX B200 node, and 5x faster queries with 83% cost reduction in production (IBM watsonx.data / Nestlé) — with no SQL changes for analysts and no Presto fork required.