Snowflake integrates with Nvidia to speed AI/ML workloads

Need for speed

AI applications, such as agents, require far more high-quality data to be accurate than traditional analytics products such as reports and dashboards. As a result, the datasets now used to inform the tools driving business decisions are significantly larger than in the past.

To process such large datasets, GPUs -- which have greater processing power than CPUs -- are used to exponentially accelerate workloads. As a result, Snowflake's native integration with Nvidia's CUDA-X libraries is a valuable addition for developers building AI tools and other advanced applications, according to Stephen Catanzano, an analyst at Informa TechTarget's Omdia.

"This integration is highly important for Snowflake users as it enables them to achieve better performance improvements," he said. "This allows data scientists to tackle computationally demanding tasks like large-scale topic modeling and computational genomics that previously would have taken hours on CPUs but can now be completed in minutes on GPUs."

This allows data scientists to tackle computationally demanding tasks like large-scale topic modeling and computational genomics that previously would have taken hours on CPUs but can now be completed in minutes on GPUs.

Michael Ni, an analyst at Constellation Research, similarly called the integration "significant" for Snowflake users given the increased development speed it enables.

"It brings GPU-class acceleration into the governed Snowflake execution environment, eliminating the handoffs that would slow ML teams down," he said. "This fills out Snowflake's offering from 'warehouse plus ML' to 'warehouse-plus-GPU compute-plus-ML platform,' joining others in the market in consolidating the data and AI platform market."

Vinay Sridhar, a product lead for AI and ML at Snowflake, noted that customers were pushing for tools that improve the performance of large-scale ML workloads. Although Snowflake was aware of customer feedback, the new integration with Nvidia was primarily motivated by the growing need for better workload performance as AI and ML development continue to rise.

OpenAI's November 2022 launch of ChatGPT represented substantial improvement in generative AI (GenAI) technology and sparked surging interest in AI development. Autonomous agents mark the latest in AI evolution, and enterprises continue increasing their investments in AI development.

"The new integration … stems from a clear technical and market need," Sridhar said. "As enterprise datasets grow exponentially, GPU acceleration is essential to maintain productivity and control costs. So, while customers were certainly pushing for faster performance, this partnership primarily evolved from Snowflake's long-term vision to enable every enterprise to achieve its full potential through data and AI."

With processing times decreasing from hours to minutes -- according to Sridhar -- Snowflake users can accelerate AI and ML workflows through the new integration while preserving existing code.

"This means teams can focus on building models and uncovering insights instead of managing complex infrastructure," Sridhar said.

The native nature of Snowflake's new integration with Nvidia, meanwhile, is unique, according to Catanzano. He also noted that many vendors have Nvidia integrations, but by embedding CUDA-X libraries within Snowflake ML and eliminating the need to change code for the libraries to work in conjunction with Snowflake tools, the integration is a potential differentiator for Snowflake.

"This is not catching up, but innovating," Catanzano said.

Ni likewise pointed out that while other vendors provide GPU acceleration capabilities, Snowflake's integration with Nvidia is unique in the way it simplifies AI and ML development by eliminating cumbersome coding rewrites.

"This is a strategic catch-up with a twist," he said. "Others have had GPU acceleration, but Snowflake … is wiring CUDA-X directly into its AI Data Cloud, delivering GPU acceleration with operational simplicity."

Snowflake's native addition of Nvidia's CUDA-X libraries is the vendor's latest move to better enable customers to build AI tools and other advanced applications.

Snowflake was slower than other vendors -- including rival Databricks and tech giants AWS, Google Cloud and Microsoft -- to react to exploding interest in AI following ChatGPT's launch. The vendor introduced some capabilities in 2023, releasing Snowflake ML as part of its Snowpark Container Services environment for AI development, but these capabilities were not extensive.

However, since Sridhar Ramaswamy was named CEO in Feb. 2024, Snowflake has made AI development a priority. With the introduction of a spate of features in June and the launch of many of them earlier this month, Snowflake meets the needs of its customers and provides AI development capabilities more in line with those of its competitors, according to Ni.

"Snowflake has flipped the script over the last year," he said. "It went from playing defense on AI to taking the offensive with a unified AI Data Cloud. By wiring Nvidia's CUDA-X stack directly into Snowflake ML and eliminating infrastructure work or code rewrites, Snowflake has moved from ‘late to the game’ to a real contender."

Catanzano also noted that Snowflake's AI development capabilities continue to improve, with the new integration furthering the vendor's evolution.

"Snowflake appears to be making substantial progress in its AI capabilities," he said.

Next steps

Looking ahead, enabling users to quickly and easily build AI tools remains Snowflake's primary product development plan, according to Sridhar.

"Snowflake is continuing to make it easier and faster for teams to build and use AI directly where their data lives," he said. "Our focus is on deepening support for our customers through AI-native integrations that simplify how data and AI workflows connect across their AI Data Cloud."

One feature that could significantly help Snowflake customers develop effective AI tools is a strong semantic layer, according to Ni. Semantic layers are governance tools that enable organizations to define their data to make it consistent and discoverable. Snowflake, in conjunction with Salesforce and a host of other vendors, is now part of a consortium developing a semantic modeling standard that simplifies the process.

"Snowflake's next step should be to strengthen the semantic and application layers that sit on top of its governed AI Data Cloud," Ni said. "With GPU-accelerated ML now running directly in Snowflake ML, the logical next step is to help customers turn that performance into repeatable, decision-ready workflows."

Catanzano, meanwhile, suggested that Snowflake not only integrate with Nvidia's data science libraries, but add integrations with AI frameworks from other vendors as well, among other possible ways the vendor could do more to serve the needs of its users.

"Snowflake could continue expanding its GPU-accelerated library ecosystem to include other specialized AI frameworks," he said. "[In addition], it could develop more automated AI model optimization tools or create industry-specific AI templates that leverage GPU capabilities to attract users who need high-performance AI processing for domain-specific use cases."

Eric Avidon is a senior news writer for Informa TechTarget and a journalist with more than 25 years of experience. He covers analytics and data management.