The public cloud was built, first and foremost, to enable companies to deliver web applications at scale. Fast forward a decade and today the cloud is used for much more than that. One of the most exciting use-cases that has emerged is leveraging the vast computational power of the cloud to run high-end workloads, such as conducting scientific experiments or training deep neural networks.
These applications have a much different usage pattern than traditional web services: They are short-lived and tend to run in batches. To respond to this new behavior, the concept of low-priority instances (commonly referred to as "spot instances") was created. Low-priority instances are essentially spare capacity in the cloud that is offered at significant discount (compared to the regular on-demand price) but with the caveat that if the capacity is needed for other tasks, they may be interrupted.
We are happy to announce that Gradient° now supports this class of instance type which we are calling "Low-Cost" instances. Low-cost instances are discounted by as much as 65%, depending on the instance type.
To run a Notebook or Job in Low-Cost mode, just add
--preemptible when using the CLI or toggle the option in the interface:
Low-cost instances function like normal instances, but differ in the following ways:
- They can interrupted at any time, even within the first few minutes.
- They are always shut down after 24 hours so they are not suitable for long-running jobs.
- They cannot be migrated to a regular VM instance.
If your workloads are fault-tolerant and can withstand possible interruptions, then Gradient Low-Cost instances are a great fit and can significantly reduce compute costs. For example, using checkpoints with TensorFlow and PyTorch will enable you to train deep learning models on Gradient Low-Cost instances without the risk of losing progress made before the instance was interrupted.
💗 The PS Engineering Team