Capacity Block for ML upgrade
I am interested in buying Capacity Blocks for ML but have one concern: If I commit to 6 or 12 month for specific instance and in the meantime something new and more powerful become available ? Will I be stuck with what i did committed 6 months ago but is there a posibility to change to a different/better performing instance ? Can customer change instance type once capacity block is bought ( assuming they will spend the same or more on new instance ) ?
- Language
- English
- Newest
- Most votes
- Most comments
When you purchase Capacity Blocks for ML, you are committing to a specific instance type for the duration of your reservation, which can be up to 6 months. Currently, there is no built-in mechanism to upgrade or change the instance type once a Capacity Block has been purchased.
The Capacity Blocks for ML service is designed to provide predictable access to specific GPU or ML chip instances for a defined period. This means that if you reserve a particular instance type, such as P5 or Trn1, you will have access to that specific type for the entire duration of your reservation.
If new, more powerful instances become available during your reservation period, you would not be able to automatically switch to these new instances within your existing Capacity Block. The reservation is tied to the specific instance type you initially selected.
It's important to note that Capacity Blocks are non-cancelable and non-modifiable once purchased. This means you cannot change the instance type, even if you're willing to pay more for a better-performing instance.
To mitigate this concern, you might consider:
- Reserving Capacity Blocks for shorter durations, which would allow you to switch to newer instances sooner if they become available.
- Carefully evaluating your long-term needs and the potential for new instance types before committing to longer reservation periods.
- Using a mix of Capacity Blocks and On-Demand instances, which would give you some flexibility to access newer instance types as they become available.
Remember, the primary benefit of Capacity Blocks is to ensure you have access to the GPU or ML chip capacity you need for your workloads at a specific time in the future. While this comes with the trade-off of being locked into a particular instance type, it provides the certainty of capacity availability for your machine learning projects.
Sources
Reserve GPU instances for ML workloads – Amazon EC2 Capacity Blocks for ML – AWS
Capacity Blocks for ML - Amazon Elastic Compute Cloud
