GPU-Accelerated Presto Delivers Low-Latency Analytics on NVIDIA GB200 NVL72
NVIDIA has demonstrated substantial latency reductions in large-scale analytical workloads by deploying GPU-accelerated Presto on next-generation hardware, including the DGX B200 and GB200 NVL72 architectures. The initiative leverages the open-source Presto distributed SQL engine integrated with NVIDIA Velox and cuDF libraries, enabling direct GPU execution for high-throughput data analytics. By utilizing NVLink 5.0 for inter-GPU communication and GPUDirect Storage for high-bandwidth data transfers, the platform minimizes host CPU overhead and system memory bottlenecks. Benchmark tests derived from TPC-H datasets revealed significant performance advantages over traditional CPU-based deployments. A single DGX B200 node equipped with eight GPUs delivered between 2.5x and 8.2x faster query runtimes compared to multi-node clusters of Intel Xeon servers processing datasets ranging from one to three terabytes. The architecture scaled effectively across multi-node environments using the GB200 NVL72 system. When tested with eight active nodes and thirty-two GPU workers connected to an IBM Storage Scale array, optimized configurations reduced overall query latency by 64 percent through targeted I/O and communication enhancements. Critical performance gains were achieved by enabling GPUDirect Storage, which routes data directly from storage to GPU memory over RDMA while maintaining NUMA node locality. Switching from POSIX reads to direct GPU reads approximately doubled performance at two-node scale. Additional optimizations included increasing I/O task sizes, utilizing dedicated threads to maximize network saturation, and implementing rebatching alongside a specific query rewrite that cut execution time for a benchmark query from fifty seconds to two seconds. These adjustments minimized compute idle time and eliminated bottlenecks associated with default exchange protocols. The GPU-accelerated Presto implementation is now integrated into the IBM watsonx.data platform, offering enterprises a technical preview for production-grade interactive dashboards and batch analytics. Developers seeking to evaluate the architecture can access nightly build containers and deployment scripts through official repositories. The demonstrated combination of high-speed interconnects, direct storage access, and GPU-native query execution establishes a scalable foundation for future data processing and low-latency analytical applications.
