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.
Speed up AI Agents.
Run SQL on NVIDIA GPUs with
Starburst Torin.
Starburst Torin is a GPU-native query engine that runs alongside our CPU engine. Your existing SQL and applications run faster on the data where it already lives, with nothing to move and nothing to rewrite.
- Up to7x
- Faster TPC-H query performance at production scale¹
- 11.4x
- Faster on the fastest TPC-H query¹
- 3.4x
- Faster throughput compared to the CPU baseline on ClickBench, at a similar hourly hardware cost¹
The problem
AI agents change the load on your data platform.
An analyst asks one question at a time. An agent working a single task can fire a chain of dependent queries, each waiting on the last. Every slow query delays the final answer.
Latency that threatens success
One agent task triggers a chain of dependent queries, so every slow query adds to the wait for the final answer.
Analytics based on CPUs
You are investing in NVIDIA GPUs for AI, while the analytical SQL that feeds those agents still runs mostly on CPUs.
Replatforming that adds time and cost
Getting more speed usually means moving data, rewriting applications, or standing up one more silo.
Why Starburst Torin
Faster analytics, using the infrastructure you already chose.
Torin adds a GPU-native engine next to Starburst's CPU engine. Run each workload on the compute that fits it, with trusted access to data wherever it lives.
Faster agents
Shorten the workload. Finish the task sooner.
Starburst Torin runs SQL natively on NVIDIA GPUs, so each step in an agent's chain returns sooner and the whole task completes sooner. The same speed lets teams run more workloads on the infrastructure they have.
- Built on NVIDIA cuDF as a new execution engine, not an acceleration layer over the CPU engine
- Scans, filters, and aggregations implemented natively for the GPU
- Inner and left joins implemented natively in the GPU engine
GPU leverage
Put your NVIDIA investment to work on analytics.
The GPUs behind your AI can also serve the SQL that feeds it. Starburst Torin extends that investment to analytics and improves price-performance where GPU execution fits.
- A GPU-native, Trino-based engine that runs alongside Starburst’s existing CPU engine
- Supported operations run on the GPU while the rest continue on the CPU, in the same query
- Runs on NVIDIA Ampere, Ada Lovelace, Hopper, and Blackwell GPUs
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.
Faster without disruption
No rewrites. No data migration. No new silo.
Speed should not require a migration project. Starburst Torin runs the SQL you already have, for the applications you already run, against data that stays where it is.
- Run existing SQL without rewriting queries
- Existing applications keep working without modification
- Query Apache Iceberg and Hive tables stored in Parquet, in place
How it works
Two engines. One SQL interface. The flexibility to run workloads on CPUs or GPUs.
You keep one SQL interface, with the flexibility to run workloads on CPUs or GPUs.
Your SQL, unchanged
Applications and agents submit the same SQL they use today. Nothing to rewrite.
Open tables, in place
Queries read Apache Iceberg and Hive tables in Parquet where they already live. No copies.
GPU where it fits
Scans, filters, aggregations, and inner and left joins run natively on the GPU.
CPU for the rest
Unsupported operations continue on the CPU engine, so the query still completes.
Built with NVIDIA
Starburst Torin,
built on NVIDIA cuDF.
Starburst built Torin on NVIDIA cuDF, with NVIDIA engineers contributing cuDF expertise and guidance on optimizing SQL operators for Blackwell GPUs.
Faster data is half the job
Agents also need access to context they can trust.
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. This work will help improve the context predictions used for fraud detection, credit risk monitoring, and know-your-customer (KYC) onboarding.
Frequently asked questions
Torin is a GPU-native query engine from Starburst. It runs alongside our CPU engine and executes supported SQL operations natively on NVIDIA GPUs, so existing queries finish faster.
No. Teams run their existing SQL and applications as they do today, without modification.
No. Torin queries Apache Iceberg and Hive tables stored in Parquet where they already live.
Table scans, filters, projections, GROUP BY and common aggregates, inner and left joins, TopN, and a set of common scalar functions. Anything else, such as right joins, DISTINCT, user-defined functions, or complex types like ARRAY and MAP, runs on the CPU engine in the same query.
It depends on the query. Gains are largest on heavy scans, joins, and aggregations over Iceberg and Hive tables, and some queries see no improvement. EXPLAIN ANALYZE shows what ran where.
An NVIDIA GPU on each Starburst Enterprise worker, with at least 24 GB of GPU memory (80 GB or more recommended). Starburst uses one GPU per worker. Production support applies to qualified hardware configurations.
Torin is in limited general availability for Starburst Enterprise, our self-managed deployment.
They continue to run on Starburst's CPU engine in the same query. Supported operations run on the GPU, and the rest run on the CPU.
Torin runs on NVIDIA Ampere, Ada Lovelace, Hopper, and Blackwell GPUs.
Torin is a new execution engine built on NVIDIA cuDF. It is not an acceleration layer over the CPU engine.
¹ Starburst 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 (NVIDIA L4), compared with an m8g.8xlarge CPU baseline. Cost figures are hardware only and exclude software licensing. Results vary by workload and deployment.
Give your agents a faster path to the answer.
Get an assessment of running Torin on your NVIDIA GPUs.