CoreWeave adds one-click Determined AI deployment for large-scale training
CoreWeave has integrated Determined AI's open-source machine learning platform, letting clients deploy large-scale training projects with one click. According to CoreWeave, projects can be up and running in seconds on its GPU infrastructure.
- CoreWeave clients can deploy Determined AI's open-source training platform with one click, according to the company.
- The integration supports multi-GPU and multi-node distributed training out of the box.
- Startups can choose between GPUs including A100s and A40s available at scale.
- CoreWeave says no GPU resources are consumed unless experiments are running.
- A demo shows training jobs ranging from 16 to 96 parallel GPUs.
What the integration offers
Determined is a deep learning training platform that combines high-performance distributed training, hyperparameter tuning, GPU scheduling and model management in a single environment. According to the company, this lets researchers and scientists focus on building models rather than managing infrastructure.
CoreWeave says its clients now have access to Determined AI resources with one click. Users integrate a small amount of code and the project is running in seconds. Determined also aims to make it simple to set up, configure, manage and share workstations or AI clusters, whether they run on-premises or in the cloud.
The benefits listed in the announcement include:
- Tracking and reproducing experiments
- Building more accurate models
- Getting more out of each GPU
- Improving developer productivity
- Speeding up time-to-market
- Improving resource utilization
- Reducing risk
Pitch to startups and cost claims
CoreWeave argues that its range and scale of GPU compute make its infrastructure a good fit for training large models, and that the integration makes deployment easier. The company says clients can get the resources they need on demand, at a fraction of the cost compared to mega-cloud providers.
The announcement points to startups in particular. Training a model is often the first step before launching a product, and CoreWeave says most large cloud providers charge a crippling amount for high-powered GPUs. Through the integration, startups can choose between GPUs, including A100s and A40s available at scale, to optimize for speed and efficiency. CoreWeave describes the result as a better performance-adjusted cost.
The company also says the platform itself consumes minimal resources once deployed. No GPU resources are used unless experiments are running, which, according to CoreWeave, prevents waste and idle compute compared with running experiments in virtual machines.
Distributed training and framework support
Determined AI on CoreWeave supports multi-GPU and multi-node distributed training out of the box. The platform supports many popular frameworks, including GPT-Neo-X, which can be used to train and fine-tune the EleutherAI 20B model.
CoreWeave also points to a demonstration by Rahul Talari, its Senior Machine Learning Engineer. He shows how to deploy Determined AI on CoreWeave and runs multiple training jobs ranging from 16 to 96 parallel GPUs.
What it means for users
For teams training large models, the integration aims to cut setup time and keep GPU spending tied to active experiments. The pricing and efficiency comparisons come from CoreWeave itself, and the announcement gives no specific prices or benchmark figures. Teams weighing the offer will need to compare costs and performance for their own workloads.
