Blog
Supercharging Scikit-Learn and Pandas in Snowflake with GPUs
Learn how to accelerate model development cycles for scikit-learn, pandas, UMAP and HDBSCAN with GPUs — no code changes required.
FEATURE
Snowflake ML
Accelerate machine learning with distributed GPUs or CPUs on the same platform as your governed data. Streamline model development and MLOps for real-time and batch workflows with no infrastructure to maintain or configure — all through a centralized UI.
Overview
Piecing together many tools for ML workflows can be complex. Get models ready for production on one platform.
Develop, deploy and monitor ML features and models with a fully integrated platform that brings together tools, real-time and batch workflows, and scalable compute infrastructure to the data.
Integrate development and MLOps
Unify model pipelines end to end with any open source model on the same platform where your data lives.
Scale models out of the box
Scale ML pipelines over CPUs or GPUs with built-in infrastructure optimizations — no manual tuning or configuration required.
Generate trusted ML insights
Discover, manage and govern features and models in Snowflake across the entire lifecycle.
ML Workflow
Accelerate development to productionwith Snowflake ML
Model Development
Build scalable models on Snowflake data with agentic ML workflows
- Autonomously generate, iterate and refine fully executable ML pipelines from natural language prompts using Cortex Code.
Optimize data loading and distribute model training from Snowflake Notebooks or any IDE of choice with ML Jobs.
- Use pre-installed libraries such as XGBoost and PyTorch, or pip install any package from open source hubs such as PyPi and HuggingFace.
Feature Management
Develop and manage features in batch and real time for production-grade pipelines
Create, manage and serve ML features with continuous, automated refresh on batch or streaming data in under 30 milliseconds using the Snowflake Feature Store.
Promote discoverability, reuse and governance features across training and inference.
Easily search for and visually trace features across the pipeline via the integrated Feature Store UI.
Production
Deploy ML models built anywhere for batch and online inference
- Log models built anywhere into Snowflake Model Registry, and serve them for batch or real-time predictions on Snowflake data with CPUs or GPUs.
Serve models in under 100 milliseconds to power low-latency, online use cases, such as personalized recommendations and fraud detection.
- Easily monitor performance and drift metrics with integrated ML Observability.
Previously, the process to train all these models and generate predictions took a half hour. The unified model on Snowflake is super quick; we’re talking minutes to generate forecasts for hundreds of thousands of customers. This speed and simplicity will help unlock additional capabilities for the business like simulation and scenario forecasting.”
Dan Shah
Manager of Data Science
“By delivering fully automated, real-time fee estimates with 97-99% accuracy in Snowflake, we are able to enhance cost transparency and strengthen confidence in settlement and reporting processes.”
Venkata Kalyan Sanaboyina
Senior Engineering Lead
Feature overview
Learn more about the integrated features for development
and production in Snowflake ML
Get Started
Take the next stepwith Snowflake
Start your 30-day free Snowflake trial today
- $400 in free usage to start
- Immediate access to the latest Snowflake ML features
- Build and deploy a model with CPUs or GPUs
End-to-end ML
FAQs
Can I use both CPUs and GPUs for ML workflows in Snowflake?
Yes, data scientists and ML engineers can build and deploy models with distributed processing in CPUs or GPUs. This is enabled by the underlying Ray-based modern container infrastructure that powers the Snowflake ML platform.
Does Snowflake support both batch and real-time ML workflows?
Yes, Snowflake ML handles both online and batch workloads. For real-time needs, our online feature store and online model inference are generally available to power use cases such as personalized recommendations, fraud detection, pricing optimization and anomaly detection.
Can I only run models in production that are built in Snowflake?
No, you can bring models built anywhere externally to run in production on Snowflake data. During inference, you can take advantage of integrated MLOps features such as ML observability and RBAC governance.
Is Snowflake ML compatible with open source?
Yes, Snowflake ML is fully compatible with any open-source library. Securely access to open source repositories via pip and bring in any model from hubs such as Hugging Face.
How does pricing work for Snowflake ML?
Snowflake operates on a consumption-based pricing model with the latest credit pricing table here.
Can I try Snowflake ML for free?
Yes, you can try any of our ML quickstarts directly from the free trial experience.
