AWS EC2 GPU instances Alternatives for Training
Training a model from scratch or continuing a pre-training run, usually for days rather than hours.
What training actually needs
Why people look past AWS EC2 GPU instances for this
The checkpoint is the cheap part to move; the built environment, the dataset cache and the exact CUDA and framework versions around it are what take a day to rebuild. That is exactly the gap AWS EC2 GPU instances does not close: If you already run on AWS, the GPU instances are next to your VPC, your IAM and your data, and that integration is worth real money. What you pay for it is the per-GPU rate and the fact that an EBS-backed environment is an AWS environment. It does not restore anywhere else.
To be fair, AWS EC2 GPU instances's real strength is real: Everything else in the account: VPC, IAM, S3 adjacency, compliance posture, committed-use and Savings Plan discounts, and a procurement path enterprises already have.
Live rates for the GPUs training wants
Live per-GPU rates from Aquanode's marketplace. Refreshes hourly.
How you run it here today
Run it today as a Pod: the console's own box workload card preselects the right template for this job.
Restoring an environment requires a snapshot that already exists. Stopping a deployment yourself captures it on the way out, so you can bring it back later on any provider. A provider-side termination is different: it is only recoverable if you had already switched automated snapshots on for that deployment, and it costs you the work since the last one. Automated snapshots are opt-in, nothing runs until you start it, and with none running there is nothing to restore.
More alternatives pages
Other workloads on AWS EC2 GPU instances
Training alternatives to other clouds
