NVIDIA GPUs Slash Presto Query Latency by 8x on GB200 NVL72

NVIDIA has demonstrated that GPU-accelerated Presto can deliver up to 8x faster query performance compared to traditional CPU clusters. Benchmarked on the NVIDIA GB200 NVL72 system, this innovation highlights the transformative potential of GPUs for large-scale data analytics, particularly for enterprises scaling their workloads across terabyte- to petabyte-sized datasets.

Presto, an open-source distributed SQL engine, is widely used for interactive queries on massive datasets. By integrating NVIDIA’s GPU technology, the platform achieves significantly reduced latency. In tests with single-node NVIDIA DGX B200 systems, Presto running on GPUs delivered 2.5x to 8.2x faster runtimes compared to CPU clusters, depending on the data scale and number of GPUs deployed. For example, using a single B200 GPU, Presto queries ran 2.5x faster than an eight-node Intel Xeon CPU cluster. With eight GPUs active, performance improved to 8.2x faster for a 1TB dataset.

The performance gains are even more pronounced when scaling to multi-node deployments on the NVIDIA GB200 NVL72 cluster. This system, built with 18 nodes featuring Grace CPUs, B200 GPUs, and high-speed NVLink connections, pairs with IBM Storage Scale for seamless data movement. Using GPU Direct Storage (GDS), data bypasses the CPU entirely, reducing overhead and accelerating query execution. For a 30TB dataset, optimized configurations on the NVL72 cluster achieved 64% faster query runtimes through I/O and communication improvements, including larger batch sizes and fine-tuning UcxExchange parameters.

GPU Acceleration: A Game Changer for Enterprise Analytics

GPU-accelerated SQL engines like Presto are rapidly gaining adoption in industries ranging from finance to retail, where interactive analytics and AI-driven insights depend on minimizing query latency. IBM and NVIDIA’s collaboration, announced at GTC 2026, has been pivotal in pushing this technology forward. Early production tests with IBM’s watsonx.data platform reported up to 25x faster query performance and 80% cost reductions compared to CPU-only setups. For enterprises, these improvements translate into faster decision-making and lower infrastructure costs.

One standout feature of NVIDIA’s GPU-accelerated Presto implementation is its use of NVIDIA cuDF libraries and NVLink 5.0 connectivity. These technologies enable high-bandwidth, low-latency GPU-to-GPU communication, making it possible to process large datasets more efficiently. Additionally, GDS optimizations ensure that data paths from storage to GPU memory avoid unnecessary CPU involvement, further enhancing throughput.

Real-World Benchmarks and Use Cases

Benchmarks based on TPC-H analytical queries demonstrate the scalability and efficiency of GPU-accelerated Presto. For a 3TB dataset, a single-node DGX B200 with three active GPUs outperformed a 10-node Intel Xeon CPU cluster by 3.6x. At the upper end, using all eight GPUs on the same node delivered 7.8x faster performance. These results underscore the suitability of GPU acceleration for compute-intensive workloads such as joins, aggregations, and scans common in data lakehouse environments.

In addition to performance gains, the integration with IBM Storage Scale adds practical value for enterprise users. The file system’s ability to handle petabyte-scale data with high I/O throughput aligns well with Presto’s requirements. During tests, switching from traditional POSIX reads to GDS-enabled reads reduced query runtimes by nearly 50%, thanks to direct data transfers from storage to GPU memory.

What’s Next for GPU-Accelerated Analytics?

The future of analytics is increasingly GPU-driven. NVIDIA and IBM are actively working to improve Presto’s GPU utilization, with plans to optimize communication between query workers and coordinators. These enhancements aim to eliminate bottlenecks and further reduce query latency. NVIDIA is also encouraging developers to test GPU-accelerated Presto through its technical preview on IBM’s watsonx.data platform.

For enterprises already investing in data lakehouse architectures, GPU acceleration offers a clear path to cost-effective and high-performance analytics. By leveraging technologies like NVIDIA GB200 NVL72, IBM Storage Scale, and GDS, businesses can achieve unprecedented query speeds, enabling faster insights and competitive advantage.

Image source: Shutterstock