Starburst Torin brings GPU-native SQL analytics to Trino

Starburst has launched Starburst Torin in limited general availability, a GPU-native analytics engine built on top of NVIDIA's cuDF library and integrated into Starburst Enterprise. The engine executes SQL workloads directly on NVIDIA GPUs, sitting alongside the platform's existing CPU engine so that enterprises can route workloads to either processor without rewriting queries, restructuring pipelines, or altering governance controls.

The company says the motivation is the distinctive query pattern introduced by AI agents. Where human analysts generate a handful of queries per session, an autonomous agent may fire dozens or hundreds of dependent queries to complete a single task. That serial chaining amplifies the effect of any per-query latency, creating a bottleneck that Starburst argues conventional CPU-based data platforms are increasingly unable to absorb at scale.

Benchmark claims

In Starburst's own TPC-H testing on AWS single-node configurations using NVIDIA Blackwell GPUs, Torin delivered a geometric-mean speedup of up to 7x across the 22-query suite, with the fastest individual query running 11.4x ahead of the CPU baseline. On ClickBench, an NVIDIA L4 GPU instance produced 3.4x the throughput of the CPU baseline at a comparable hourly hardware cost. The benchmarks were run in development environments on Starburst Enterprise release 483-e; the company notes that results vary by workload and deployment.

Justin Borgman, co-founder and chief executive at Starburst, said: "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."

Todd Mostak, senior director of analytics and data intelligence at NVIDIA, added that Torin allows customers to run existing SQL on NVIDIA GPUs "without moving data, rebuilding workflows, or changing governance controls." The two companies co-developed the engine, with NVIDIA engineers contributing cuDF expertise and guidance on tuning for Blackwell architecture. Separately, Starburst's CPU engine has been validated on NVIDIA's Vera processor, which in Starburst's TPC-H testing ran 1.5x faster than a leading x86 instance while using approximately 33% fewer CPU-seconds per query.

Starburst is also working to integrate NVIDIA's open Nemotron Retriever and Kumo models into its AIDA product, targeting natural-language analytics for regulated financial-services customers. Those models are part of NVIDIA's Agent Knowledge Foundation and Agent Toolkit, aimed at helping AI agents retrieve grounded context and generate reliable answers for use cases such as fraud detection, credit risk monitoring, and know-your-customer onboarding.

Market context

The GPU-accelerated analytics space is crowded and moving quickly. DuckDB, Apache Arrow and GPU-native query engines from vendors including Voltron Data and RAPIDS have been expanding the ecosystem of cuDF-compatible tooling over the past two years. Hyperscalers are also integrating accelerated compute into their managed data warehousing products, raising the baseline against which specialist platforms must differentiate.

Starburst's positioning is federation-first: the platform is designed to query data in place across on-premises systems, multiple clouds, and hybrid architectures, underpinned by Trino and Apache Iceberg. That architecture is a deliberate hedge against the lock-in risk that comes with moving data into a single proprietary store. The Torin announcement extends that value proposition to GPU compute, allowing enterprises already investing in NVIDIA AI infrastructure to apply those assets to SQL analytics without adopting a second data platform.

Starburst counts Comcast, Citigroup, and four of the top five global banks among its customers, providing commercial credibility as it pitches Torin to the financial-services sector. Named first Torin customer deployments and production-grade pricing have not yet been disclosed; both will be closely watched metrics as the product moves from limited to full general availability.