NVIDIA Speeds Up Presto Analytics Queries on GB200 NVL72 GPU Platform

In a post for developers, NVIDIA explained how GPU acceleration of the open source Presto SQL engine on the GB200 NVL72 platform lets it run analytics queries against very large datasets with low latency.

What is Presto and why does it matter?

Presto is an open source distributed SQL engine originally built for fast interactive queries over massive datasets, without first moving that data into a separate analytics warehouse. Analysts and data engineers turn to it when they need an answer to a SQL query in seconds, not after the lengthy batch processing typical of classic big data tools.

What GPU acceleration delivers

According to NVIDIA, when Presto's compute moves onto GB200 NVL72 GPU infrastructure, the engine hits peak performance specifically on low-latency analytics workloads — scenarios where response time matters while working with very large volumes of data. This isn't about replacing Presto with a new product. The goal is for the same open source SQL engine to make better use of GPUs instead of conventional CPU clusters.

The gap between CPU and GPU execution of analytics queries is most visible under high concurrency, when many analysts hit the same huge dataset at once and expect answers in real time rather than after a long queue of batch jobs. These are exactly the low-latency scenarios NVIDIA is targeting with the Presto and GB200 NVL72 pairing.

• Presto — an open source distributed SQL engine for interactive queries over big data
• Acceleration platform — NVIDIA GB200 NVL72, the company's GPU infrastructure
• Claimed effect — peak performance specifically on low-latency analytics workloads
• The post was published on the NVIDIA Developer Blog for developers and data engineers

Why NVIDIA is pushing GPUs for analytics

The post continues NVIDIA's push to position GPUs not only for training and inference of AI models but also for classic database analytics. The company is encouraging developers to move existing SQL engines such as Presto onto GPU platforms like GB200 NVL72, rather than relying solely on scaling out CPU clusters for interactive analytics. For teams already running AI workloads on GB200 NVL72, this means they can use the same infrastructure for analytics SQL queries without standing up a separate CPU environment for data.

The approach makes economic sense for organizations that have already invested in GPU clusters for AI training and inference: instead of maintaining separate CPU infrastructure purely for SQL analytics, they can spread analytics and AI workloads across the same fleet of GPU servers.

What it means

The growth of GPU-accelerated versions of popular open source tools like Presto shows that the competition over data infrastructure now extends beyond neural network training. NVIDIA increasingly positions GPUs as a general-purpose compute platform not just for AI but also for everyday business analytics over very large datasets. For the industry as a whole, it's another sign that the line between AI infrastructure and data infrastructure keeps blurring, and that GPU purchasing decisions increasingly factor in not only model training but also the everyday SQL workloads of analytics teams.

Who benefits most from the pairing?

The teams that gain the most from this acceleration are those whose big data analytics queries already run into CPU infrastructure latency limits, and whose data is too large or too volatile to prepare in advance in a separate warehouse for each specific report.