In today's data-driven world, the demand for rapid data processing is ever-increasing. Businesses are constantly seeking ways to accelerate their analytics pipelines to gain real-time insights and make informed decisions. GPU-accelerated databases and query engines have emerged as a powerful solution, offering substantial price-performance advantages over traditional CPU-based systems. The parallel processing capabilities of GPUs, with their high memory bandwidth and numerous threads, are particularly well-suited for compute-intensive tasks such as complex joins, aggregations, and string processing. Now, IBM and NVIDIA are joining forces to bring NVIDIA cuDF to the Velox execution engine, unlocking GPU-native query execution for widely used platforms like Presto and Apache Spark. This collaboration promises to revolutionize large-scale data analytics, but only if businesses understand how to leverage this technology effectively.

The collaboration between IBM and NVIDIA introduces several key advancements:

This collaboration has significant implications for organizations dealing with large-scale data analytics. By leveraging GPU acceleration, businesses can achieve:

The performance gains achieved through this collaboration are substantial. For example, in Presto tpch benchmarks at scale factor 1,000, Presto on NVIDIA GH200 Grace Hopper Superchip achieved a runtime of 99.9 seconds for 21 of 22 queries, compared to 1,246 seconds for Presto C++ on AMD 7965WX. Multi-GPU execution on an eight-GPU NVIDIA DGX A100 node delivered over 6x speedup using NVLink in the exchange operator compared to the Presto baseline HTTP exchange. In Apache Spark, offloading the compute-intensive second stage of TPC-DS Query 95 SF100 to GPU resulted in significant performance improvements, even when the first stage was executed on CPU. The table below illustrates the performance comparison:

| Benchmark | Platform | Runtime (seconds) | Queries Completed | Notes | | ---------------------------------------- | ------------------------------------------- | ----------------- | ----------------- | --------------------------------------- | | Presto tpch (SF1000) | Presto C++ on AMD 7965WX | 1,246 | 21/22 | CPU-only | | Presto tpch (SF1000) | Presto on NVIDIA RTX PRO 6000 Blackwell | 133.8 | 21/22 | Single GPU | | Presto tpch (SF1000) | Presto on NVIDIA GH200 Grace Hopper Superchip | 99.9 | 21/22 | Single GPU | | Presto tpch (SF1000) | Presto GPU on NVIDIA GH200 | 148.9 | 22/22 | Single GPU, CUDA managed memory | | Presto tpch (SF1000) | Presto GPU on NVIDIA DGX A100 (8 GPUs) | N/A | 22/22 | Multi-GPU, UCX-based Exchange, NVLink | | Gluten tpcds Query 95 SF100 (Stage 2 GPU) | CPU (8 vCPUs) + NVIDIA T4 GPU (g4dn.2xlarge) | Faster | N/A | Hybrid CPU-GPU, Stage 1 CPU, Stage 2 GPU |

Velox supports several Exchange types for different types of data movements: Partitioned, Merge, and Broadcast. Partitioned Exchange uses a hash function to partition input data. Merge Exchange receives multiple input partitions from other workers and then produces a single, sorted output partition. Broadcast Exchange loads the data in one worker and then copies the data to all other workers.

The collaboration between IBM and NVIDIA to bring GPU-native Velox and NVIDIA cuDF to Presto and Apache Spark represents a significant leap forward in large-scale data analytics. By unlocking the power of GPUs, organizations can achieve unprecedented performance gains, enabling faster insights and improved decision-making. As the data landscape continues to evolve, GPU acceleration will become increasingly critical for remaining competitive. We anticipate further advancements in this area, with wider adoption of GPU-powered data analytics across various industries. This is a space to watch closely as it shapes the future of data processing.

Sources verified via NVIDIA of October 6, 2025.