Starburst Introduces Torin, a GPU-Native Analytics Engine Built on NVIDIA cuDF

Starburst Torin runs SQL natively on NVIDIA GPUs, delivering up to 7x faster query performance to accelerate AI agents and enable data teams to run more workloads without moving data or changing existing SQL or applications.

BOSTON, October 21, 2026 - Starburst today announced limited general availability of Starburst Torin, a GPU-native, Trino-based analytics engine for Starburst Enterprise. Built on NVIDIA cuDF, Torin accelerates SQL workloads with native GPU execution, delivering up to 7x faster query performance.1 By complementing Starburst's existing CPU engine, Torin gives enterprises the flexibility to run workloads on CPUs or GPUs without changing existing Trino SQL queries, applications, or agents.

AI agents create a different kind of load on data platforms. A single agent task can trigger a chain of dependent queries, where delays in one query can slow the entire workflow. These bottlenecks also affect interactive analytics, making it harder to explore data, evaluate results, and make decisions in real time. Faster query execution helps agents complete analytical work sooner, improves user responsiveness, and lets teams run more workloads on available infrastructure.

Meanwhile, enterprises are investing in NVIDIA AI infrastructure to accelerate data processing for the agentic era. Torin extends those investments to SQL analytics without requiring a new data platform.

"Speed matters because of what it unlocks, which is why we built Starburst Torin. AI agents run dozens or hundreds of queries to finish a single task, so faster queries mean agents complete more work, teams run more workloads, and the business gets results sooner," said Justin Borgman, Co-Founder and CEO at Starburst.

In Starburst benchmarks, Torin ran TPC-H queries up to 7x faster on NVIDIA Blackwell GPUs, with the fastest query running 11.4x faster. On ClickBench, an NVIDIA L4 GPU instance delivered 3.4x the throughput of the CPU baseline at a similar hourly hardware cost.1

Torin is the result of close engineering work between Starburst and NVIDIA. Starburst built Torin's SQL operators on NVIDIA cuDF, and NVIDIA engineers contributed cuDF expertise and guidance on optimizing for NVIDIA Blackwell GPUs. The results bring NVIDIA's accelerated computing to the SQL teams already using Starburst. On the CPU side, Starburst Enterprise's CPU engine also supports the NVIDIA Vera CPU, announced at GTC in March. For Starburst's TPC-H benchmark at SF1000, Vera ran the queries 1.5x faster than a leading x86 CPU and used about 33% fewer CPU seconds per query.2

"Analytics and AI agents are putting more demand on data platforms, and accelerated data processing is how they keep up. By building Starburst Torin on NVIDIA cuDF, Starburst lets customers run their existing SQL on NVIDIA GPUs, without moving data, rebuilding workflows, or changing governance controls," said Todd Mostak, Senior Director of Analytics and Data Intelligence at NVIDIA.

Fast queries are only half of what agents need. They also need context they can trust, and Starburst is working with NVIDIA to address that need. To deliver reliable natural-language business insights to highly regulated financial services customers, Starburst is working to integrate open NVIDIA Nemotron Retriever and NVIDIA Kumo models for inclusion in Starburst AIDA. Both are part of the new NVIDIA Agent Knowledge Foundation, which provides open models and accelerated libraries within the NVIDIA Agent Toolkit, helping agents retrieve relevant context, generate grounded answers, analyze and predict outcomes, and optimize actions. This work will help improve the context predictions used for fraud detection, credit risk monitoring, and know-your-customer (KYC) onboarding.

Starburst is at GTC Berlin at booth 3072. Justin Borgman joins a panel on accelerated data processing for AI workloads on Thursday, October 22, at 4 p.m. CEST.

About Starburst

Starburst delivers enterprise intelligence at scale by giving organizations secure, governed access to all their data, wherever it lives. Built for distributed data environments across on-premises systems, multiple clouds, and hybrid architectures, the Starburst platform helps companies power AI and analytics without the cost, complexity, and limitations of traditional data consolidation.

By enabling federated access to complete enterprise context, Starburst helps AI platforms, applications, and teams move from pilots to production with trusted, actionable data. Business leaders gain faster insights, stronger collaboration, and greater confidence in decisions based on complete and governed data. Built on open standards with Trino and Apache Iceberg, Starburst integrates with existing technologies while helping organizations avoid vendor lock-in.

Leading enterprises in more than 60 countries, including Comcast, Citigroup, and four of the top five global banks, trust Starburst to fuel enterprise intelligence. For more information, visit starburst.ai.

Related Resources:
Read the blog from Martin Traverso, Starburst CTO and co-creator of Trino: AI Agents Are Outgrowing the Converged Data Platform

Starburst and NVIDIA Overview

¹ The TPC-H and ClickBench results in this release are from the testing published in the Starburst engineering blog. Up to 7x reflects the same TPC-H testing on the same NVIDIA Blackwell GPUs, run on Starburst Enterprise 483-e. Benchmarks ran on AWS in a development environment on single-node configurations, with warmup runs followed by measured runs. Speedups are geometric means across per-query ratios. TPC-H ran on g7e.4xlarge instances and ClickBench on g6.4xlarge instances, compared with an m8g.8xlarge CPU baseline. Cost figures are hardware only and exclude software licensing. Results vary by workload and deployment. Source: Starburst engineering blog.

2 Benchmarks ran in a development environment on single-node configurations with 88 physical cores / 176 threads for NVIDIA Vera and 96 physical cores / 192 threads for Intel Xeon 6, on Starburst Enterprise 483-e, one query at a time. Workloads were TPC-H (22 queries). NVIDIA Vera (single socket) ran 1.48x faster than an Intel Xeon 6 instance, measured by query wall time, and used about 33% fewer CPU seconds per query. Results vary by workload and deployment.