Starburst supports GPUs for faster distributed data analysis

retained text

Starburst supports GPUs for faster distributed data analysis
Starburst is getting 3 to 4x faster responses to AI analysis queries by using GPUs instead of CPUs.
The company develops and uses Trino open source distributed SQL to query and analyze distributed data sources, and uses x86 CPUs to process queries. Its software engine enables data querying across spreadsheets, databases, cloud warehouses and lakehouses. In May, Dell updated its AI Factory product set and, deep in the details was this: the platform has a Starburst-powered Data Analytics Engine with GPU-accelerated SQL analytics that delivers up to 6x faster query performance than non-GPU-accelerated systems on Nvidia Blackwell GPUs today with support designed for future Vera systems.”
A briefing from Starburst CEO Justin Borgman revealed that Starburst’s software now supports Nvidia’s Vera CPU and Rubin GPU. We got there in a roundabout way by asking how is Starburst positioned vis-a-vis Databricks and Snowflake?
Borgman told us: “It is classic co-opetition, right? We cooperate in the sense from a customer perspective, we inter-operate with those platforms. We query those platforms. So if you're a large bank and you have a little Databricks, a little Snowflake, you probably have also Teradata and IBM and Oracle and all these other things, Starburst is very complimentary to everything you have. And that's an important part of our strategy to provide these customers with the optionality to work with their existing investments. The competition side, of course, is that those two platforms want to be the single source.”
So “We do compete for workloads here and there, but we have a different architectural approach. And I think for large enterprise, we have a very particular value proposition that those two can't match, which is both the federation and hybrid deployment. We can work on-prem, those two are cloud-only.”
And then this: “That’s where the partnership with Nvidia has become very interesting because we now also support GPUs in addition to CPUs. And Nvidia has also now a new CPU they've launched this year, but of course GPUs are most of their business and you can now use existing GPUs that you bought, maybe you have some idle capacity, you can run analytics workloads now leveraging those GPUs as well. More flexibility.”
He added: “We've been working on that from the beginning of this year and we are just releasing now our first version of it and it uses something that NVIDIA built called cuDF .” This, Nvidia says, is an open source NVIDIA CUDA-X data processing toolkit for structured data that delivers massive speedups and cost savings for data engines and libraries. Built on highly optimized Nvidia CUDA primitives, cuDF taps into GPU parallelism and memory bandwidth to accelerate data processing and analytics workflows.
Borgman said: “We're seeing really good performance, actually about three to four times faster than what you'd see on a CPU. Now, GPUs are a lot more expensive, so that is still something you have to factor in from a cost-performance perspective. But what we found as an early use case is a lot of these big banks, … they’ve all bought GPUs and they're not using them all the time at 100 percent. So they all have idle capacity that they've already paid for, right? It's CapEx that they paid upfront and the ability to maybe increase utilisation from 60 to 90 percent by running some batch analytic workloads in there overnight or when it's idle.”
He mentioned two ways of doing this: “One is just if you know every night from nine to midnight , you've got a regular [free slot] you can have it scheduled, but what we've also done is, we can measure the GPU utilisation and then, based on that, decide whether to slot in some jobs.”
Nvidia likes this “because for them it’s TAM expansion for a $5 trillion company, if you will.”
We asked Borgman what triggered this idea?
He said: “It's funny how much serendipity plays in decisions or creating opportunities. For us at the beginning of the year, Nvidia reached out to us to benchmark on their new CPU. So they have this CPU called the Vera processor, right? And this is of course their entrance into the CPU market. And so we benchmarked it and it turned out we were really fast on Vera. In fact, much faster on Vera than an Intel processor. So they liked that. They put that in Jensen's keynote and promoted our benchmarks. That began the relationship, at which point they were like, "Hey, we have this cuDF framework that allows you to run things on GPUs. Would you mind taking a look at that and see if you can make Starburst run on GPUs?”"
The result was a speedup “anywhere from three to seven X, depending on the query type,” with the 7X gain on the bigger, more complex queries.
This, Borgman suggests, is going to affect the Starburst and Databricks/Snowflake market positioning: “We think this is a unique position for us because Databricks and Snowflake, as you know, they’re cloud-hosted offerings. They have to carry the cost of processing within their cost structure. It impacts their margin. We don't. We allow you to deploy it on your own infrastructure. So naturally, Databricks and Snowflake are always going to be looking for what is the cheapest compute they can find, which doesn't lend itself very well to Nvidia necessarily. Nvidia is not the cheapest and doesn't want to be the cheapest. So we think we're really the only data platform that can be a great partner for Nvidia here in this enterprise market.”

document fetched 2026-09-19 02:57 · canonical https://blocksandfiles.com/architecture/2026/09/18/starburst-supports-gpus-for-faster-distributed-data-analysis/5297256