NVIDIA GB200 NVL72 accelerates Presto queries by 8x for streaming telemetry
NVIDIA released technical benchmarks showing that its GB200 NVL72 and DGX B200 hardware significantly accelerate Presto SQL query performance compared to traditional CPU-based configurations. The company emphasizes the use of GPU-Direct Storage and NVLink technology to reduce latency for large-scale data analytics relevant to streaming platform telemetry and viewer data management.
Key Takeaways
- GPU-accelerated Presto on a single DGX B200 node delivered 8.2x lower latency than an eight-node Intel Xeon CPU cluster.
- NVIDIA GPUDirect Storage (GDS) with IBM Storage Scale achieved 2x faster runtimes by bypassing host CPU bounce buffers.
- The GB200 NVL72 configuration integrates 72 B200 GPUs connected via NVLink, providing 1,800 GB/s of bidirectional bandwidth.
- Software optimizations, including 16 MiB I/O tasks and query rewrites for GPU-based Parquet writing, improved total runtime by 64%.
Why It Matters
The transition from CPU to GPU-accelerated analytics addresses the ballooning telemetry demands of global streaming services. As platforms shift from subscriber growth to revenue optimization, the ability to process petabytes of viewer data with sub-second latency is critical for dynamic ad insertion and churn prediction. This hardware stack concretely reduces the infrastructure footprint by replacing dozen-node CPU clusters with single-rack GPU systems, lowering TCO while improving data freshness. We should watch for the integration of this GPU-native Presto into broad managed services like IBM watsonx.data to see if it becomes the standard for large-scale AVOD performance monitoring.
Additional Context
The demand for high-speed telemetry processing comes as global online video subscriptions reached 2.24 billion in 2025, per Omdia in June 2026. This massive scale has shifted industry focus toward maximizing revenue from existing users through data-driven personalization. According to Precedence Research in May 2026, the streaming market is projected to reach $195.85 billion this year, with ad-supported tiers (AVOD) and FAST models growing at a 14.7% CAGR. These models require massive, low-latency analytical backends to manage real-time ad bidding and delivery signals across billions of monthly active viewers. NVIDIA's push into rack-scale analytics with the GB200 NVL72 occurs amidst a broader hardware transition. Per Silicon Analysts in March 2026, the B200 GPU's manufacturing cost is roughly double that of previous generations, driven by an 84% gross margin and the high cost of HBM3e memory. To justify these capital expenditures, enterprises are seeking 1/15th relative cost-per-token efficiencies in inference and analytics. NVIDIA's "co-design" approach—bundling GPUs with NVLink and specialized storage protocols—is designed to lock in hyperscalers who are otherwise exploring custom silicon like Google's TPU or AWS Trainium, per reports from April 2026. Furthermore, the partnership with IBM to certify Storage Scale 6.0 for NVIDIA SuperPOD architectures, noted in June 2026, addresses a critical bottleneck: nearly 40% of large-scale data projects fail due to I/O starvation. By bypassing the CPU, the GB200 architecture allows streaming engineers to query 10PB+ data lakes—a scale first popularized by Netflix's use of Presto—without the latency penalties of traditional POSIX-based file systems. This convergence of high-capacity storage and liquid-cooled compute racks, which now reach 140 kW per rack, represents the new baseline for top-tier streaming infrastructure.
Read full article at developer.nvidia.com
Enjoy our coverage?
Add StreamingMeme as a preferred source on Google to see more of our streaming news at the top of your Search results.
Add as preferred source