Innovating with AI requires accelerators such as GPUs that can be hard to come by in times of extreme demand. To address this challenge, we offer Dynamic Workload Scheduler (DWS), a service that optimizes access to compute resources when and where you need them. In July, we announced Calendar mode in DWS to provide short-term ML capacity without long-term commitments, and today, we are taking the next step: the general availability (GA) of Flex-start VMs.
Available through the Compute Engine instance API, gcloud CLI, and the Google Cloud console, Flex-start VMs provide a simple and direct way to create single VM instances that can wait for in-demand GPUs. This makes it easy to integrate this flexible consumption option into your existing workflows and schedulers.
What are Flex-start VMs?
Flex-start VMs, powered by Dynamic Workload Scheduler, introduce a highly differentiated consumption model that’s a first among major cloud providers, letting you create single VM instances that provide fair and improved access to GPUs. Flex-start VMs are ideal for defined-duration tasks such as AI model fine-tuning, batch inference, HPC, and research experiments that don’t need to start immediately. In exchange for being flexible with start time, you get two major benefits:
- Dramatically improved resource obtainability: By allowing your capacity requests to persist in a queue for up to two hours, you increase the likelihood of securing resources, without needing to build your own retry logic.
- Cost-effective pricing: Flex-start VM SKUs offer significant discounts compared to standard on-demand pricing, making cutting-edge accelerators more accessible.
Flex-start VMs can run uninterrupted for a maximum of seven days and consume preemptible quota.
A new way to request capacity
Source Credit: https://cloud.google.com/blog/products/compute/introducing-flex-start-vms-for-the-compute-engine-instance-api/
