With Amazon EC2 Capacity Blocks for ML, AWS aims to lower the barrier for companies that need GPU computing power to experiment with AI.

Amazon Web Services introduces Amazon EC2 Capacity Blocks for ML. This offering allows GPU capacity from an EC2 Ultracluster to be reserved in advance. AWS compares the system to that of hotel reservations: instead of booking a room of a specific type with an arrival and departure date, you reserve a lot of GPU capacity with a start and end date. Once your reservation begins, you will have access to the capacity and can launch EC2 P5 instances that use it. At the end of the reservation, any running instances will be shut down.

Dynamic price

Like a hotel, AWS wants to dynamically price GPU capacity based on demand. The system guarantees the cloud provider that the significant investments in GPU computing power are used to the maximum. Users are assured that they can use a certain amount of GPU processing power for a given budget, although the offer appears to be of particular interest to AWS. Like other major cloud providers, they have a difficult path to navigate: AU hype is driving demand for GPU instances, but GPUs themselves are a scarce resource that doesn’t roll off the assembly line fast enough. Capacity blocks for ML are one way to deal with this.

The offer is available now, but only in AWS US East (Ohio). Anyone who signs up will have access to up to 64 instances with eight Nvidia H100 Tensor Core GPUs for up to fourteen days. The approach is interesting because it doesn’t really correspond to what cloud providers usually do: flexibility and pay-as-you-go are replaced by reservation and prepayment.