Running Low-Latency Analytical Workloads with GPU-Accelerated Presto on NVIDIA GB200 NVL72

2026-07-08 · Source: NVIDIA Technical Blog · Field: Technology & Digital — Data Science & Analytics, Cloud Computing & IT Infrastructure, Artificial Intelligence & Machine Learning · Depth: Advanced, medium

Summary

GPU-accelerated Presto, an open-source distributed SQL engine, delivers peak performance and low latency for analytical query workloads by leveraging NVIDIA GPUs. This technology utilizes NVIDIA cuDF algorithms and NVIDIA NVLink for fast GPU-to-GPU communication. Benchmarks on a single-node NVIDIA DGX B200 demonstrated 2.5x to 8.2x faster runtimes for TPC-H-derived datasets (1K and 3K scale factors) compared to 8-10 node Intel Xeon 6642Y CPU clusters. The system also scales out efficiently on multi-node NVIDIA GB200 NVL72 clusters, achieving 64% faster query runtimes through I/O and communication optimizations, including NVIDIA GPUDirect Storage (GDS) and IBM Storage Scale. GDS, which bypasses host CPU and system memory, showed approximately 2x faster runtimes than POSIX cold reads for 10K scale factor datasets on NVL72, making it the preferred approach for performance and cost-efficiency. GPU-accelerated Presto is now available for testing via a technical preview on the IBM watsonx.data platform.

Key takeaway

For Data Engineers or MLOps Engineers deploying large-scale analytical SQL workloads, you should evaluate GPU-accelerated Presto on NVIDIA hardware. This solution offers up to 8x faster query runtimes and 64% overall performance improvements by leveraging NVIDIA cuDF, NVLink, and GPUDirect Storage. Consider testing the technical preview on IBM watsonx.data or exploring the Presto Native "gpu-nightly" Docker image to validate its low-latency and high-throughput capabilities for your production environments.

Key insights

GPU-accelerated Presto delivers substantial latency reduction for analytical workloads through optimized hardware and software integration.

Principles

• GPU-native libraries and high-speed interconnects are vital for query acceleration.
• Direct storage-to-GPU data paths significantly reduce I/O bottlenecks.
• I/O task sizing and thread count impact NVLink saturation and overall runtime.

Method

Benchmarking involved TPC-H derived queries, parquet files, and averaging runtimes; optimizations included GDS, 16 MiB I/O tasks, 16 I/O threads, rebatching, and query rewrites.

In practice

• Deploy GPU-accelerated Presto with NVIDIA cuDF and NVLink for performance.
• Integrate NVIDIA GDS with IBM Storage Scale for direct data transfer to GPU memory.
• Tune I/O parameters like 16 MiB task size and 16 I/O threads for optimal throughput.

Topics

• GPU-accelerated Presto
• NVIDIA GB200 NVL72
• GPUDirect Storage
• Analytical Workloads
• TPC-H Benchmarking
• IBM Storage Scale

Code references

• rapidsai/cudf
• rapidsai/velox-testing
• prestodb/prestorials

Best for: CTO, VP of Engineering/Data, Director of AI/ML, Data Engineer, MLOps Engineer, AI Architect

Related on AIssential

• Designing GPU-Accelerated Query Engines with NVIDIA GQE — GPU-accelerated query engines achieve significant speedups by optimizing data movement, compression, and execution on NVIDIA hardware.
• Using Accelerated Computing to Live-Steer Scientific Experiments at Massive Research Facilities — GPU-accelerated computing enables real-time analysis and steering of exascale scientific experiments, drastically reducing data processing times.
• Fixing GPU Starvation in Large-Scale Distributed Training — GPU underutilization often stems from data I/O and format transformation bottlenecks, not model complexity.
• Accelerate LLM model loading and increase context windows with GPUDirect on Amazon FSx for Lustre and TurboQuant — Direct storage-to-GPU data paths and KV cache compression drastically cut LLM load times and expand context windows.
• Lambda's MLPerf Inference v6.0: hardware leap, software maturity, research breakthrough — Hardware, software, and routing optimizations significantly boost AI inference performance and reduce latency.
• GPU Scaling from 1 to 261120 threads Part 1 — GPU performance scales non-linearly with threads due to memory latency and inefficient warp utilization.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by NVIDIA Technical Blog.