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

NVIDIA Reports Up to 8x Lower Latency with GPU-Accelerated Presto

NVIDIA has released new performance benchmarks demonstrating that its GPU-accelerated Presto SQL engine, running on DGX B200 and GB200 NVL72 systems, can achieve up to eight times lower latency on analytical workloads compared to traditional multi-node CPU clusters. The findings are significant for enterprises running large-scale data analytics, as reducing query time is critical for interactive dashboards and the operational efficiency of AI agents that rely on fast data retrieval. By consolidating workloads onto fewer, more powerful GPU-based nodes, the architecture presents a new approach to handling terabyte-scale datasets efficiently.

The performance gains stem from a tightly integrated hardware and software stack. At its core, the system uses NVIDIA's cuDF library for query execution, NVLink for high-speed interconnect between GPUs, and GPUDirect Storage (GDS) for high-throughput data transfers directly from storage to GPU memory. When paired with an IBM Storage Scale system, GDS was shown to be 2x faster than conventional POSIX reads by bypassing CPU bounce buffers and avoiding NUMA-related penalties. Further cluster-level optimizations on the GB200 NVL72 yielded a cumulative 64% improvement in query runtimes through a series of specific tunings:

• Switching to device reads with GDS and increasing I/O task size to 16 MiB delivered a ~30% speedup.
• Increasing I/O threads from four to 16 resulted in an additional ~17% faster runtime by better saturating the NVLink fabric.
• Applying query-specific rewrites and using larger exchange batch sizes reduced the final runtime by another 35%.

These results signal a potential shift in data infrastructure strategy for the broader AI market. The ability of a single DGX B200 node with eight GPUs to outperform a 10-node CPU cluster challenges the established economic and operational models of building scale-out data analytics platforms. With GPU-accelerated Presto now entering technical preview on IBM's watsonx.data platform, the technology is moving from benchmark to production viability. This development pressures organizations to re-evaluate the total cost of ownership (TCO) for their data stacks, weighing the capital expense of advanced GPU systems against the operational savings from reduced hardware footprint, power consumption, and faster time-to-insight.

NVIDIA's results demonstrate a significant shift in data analytics architecture, where tight integration between GPU hardware (GB200), direct storage access (GDS), and optimized software (cuDF) can deliver superior performance-per-dollar compared to traditional scale-out CPU clusters, effectively redefining the TCO for high-throughput SQL query engines.