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

NVIDIA has demonstrated that GPU-accelerated Presto can ship as much as 8x sooner question efficiency in comparison with conventional CPU clusters. Benchmarked on the NVIDIA GB200 NVL72 system, this innovation highlights the transformative potential of GPUs for large-scale information analytics, notably for enterprises scaling their workloads throughout terabyte- to petabyte-sized datasets.

Presto, an open-source distributed SQL engine, is broadly used for interactive queries on huge datasets. By integrating NVIDIA’s GPU know-how, the platform achieves considerably diminished latency. In assessments with single-node NVIDIA DGX B200 methods, Presto working on GPUs delivered 2.5x to eight.2x sooner runtimes in comparison with CPU clusters, relying on the info scale and variety of GPUs deployed. For instance, utilizing a single B200 GPU, Presto queries ran 2.5x sooner than an eight-node Intel Xeon CPU cluster. With eight GPUs lively, efficiency improved to eight.2x sooner for a 1TB dataset.

The efficiency positive aspects are much more pronounced when scaling to multi-node deployments on the NVIDIA GB200 NVL72 cluster. This method, constructed with 18 nodes that includes Grace CPUs, B200 GPUs, and high-speed NVLink connections, pairs with IBM Storage Scale for seamless information motion. Utilizing GPU Direct Storage (GDS), information bypasses the CPU totally, lowering overhead and accelerating question execution. For a 30TB dataset, optimized configurations on the NVL72 cluster achieved 64% sooner question runtimes by I/O and communication enhancements, together with bigger batch sizes and fine-tuning UcxExchange parameters.

GPU Acceleration: A Recreation Changer for Enterprise Analytics

GPU-accelerated SQL engines like Presto are quickly gaining adoption in industries starting from finance to retail, the place interactive analytics and AI-driven insights rely on minimizing question latency. IBM and NVIDIA’s collaboration, introduced at GTC 2026, has been pivotal in pushing this know-how ahead. Early manufacturing assessments with IBM’s watsonx.information platform reported as much as 25x sooner question efficiency and 80% value reductions in comparison with CPU-only setups. For enterprises, these enhancements translate into sooner decision-making and decrease infrastructure prices.

One standout function of NVIDIA’s GPU-accelerated Presto implementation is its use of NVIDIA cuDF libraries and NVLink 5.0 connectivity. These applied sciences allow high-bandwidth, low-latency GPU-to-GPU communication, making it attainable to course of massive datasets extra effectively. Moreover, GDS optimizations be sure that information paths from storage to GPU reminiscence keep away from pointless CPU involvement, additional enhancing throughput.

Actual-World Benchmarks and Use Circumstances

Benchmarks based mostly on TPC-H analytical queries show the scalability and effectivity of GPU-accelerated Presto. For a 3TB dataset, a single-node DGX B200 with three lively GPUs outperformed a 10-node Intel Xeon CPU cluster by 3.6x. On the higher finish, utilizing all eight GPUs on the identical node delivered 7.8x sooner efficiency. These outcomes underscore the suitability of GPU acceleration for compute-intensive workloads similar to joins, aggregations, and scans frequent in information lakehouse environments.

Along with efficiency positive aspects, the mixing with IBM Storage Scale provides sensible worth for enterprise customers. The file system’s means to deal with petabyte-scale information with excessive I/O throughput aligns nicely with Presto’s necessities. Throughout assessments, switching from conventional POSIX reads to GDS-enabled reads diminished question runtimes by practically 50%, due to direct information transfers from storage to GPU reminiscence.

What’s Subsequent for GPU-Accelerated Analytics?

The way forward for analytics is more and more GPU-driven. NVIDIA and IBM are actively working to enhance Presto’s GPU utilization, with plans to optimize communication between question staff and coordinators. These enhancements intention to get rid of bottlenecks and additional scale back question latency. NVIDIA can also be encouraging builders to check GPU-accelerated Presto by its technical preview on IBM’s watsonx.information platform.

For enterprises already investing in information lakehouse architectures, GPU acceleration affords a transparent path to cost-effective and high-performance analytics. By leveraging applied sciences like NVIDIA GB200 NVL72, IBM Storage Scale, and GDS, companies can obtain unprecedented question speeds, enabling sooner insights and aggressive benefit.

Picture supply: Shutterstock