AWS slices L4 GPUs into eight ECS workloads

Amazon ECS can now schedule a container on one-eighth of an Nvidia L4 GPU, giving each smallest partition 3 GB of memory on EC2 G6f instances. AWS’s fractional-GPU release exposes task-definition requests of 0.125, 0.25, or 0.5 GPU across ECS Managed Instances and ECS on EC2. Builders sizing small-model inference should treat the feature as a utilization lever beside AMD’s argument for a broader accelerator shortlist: benchmark isolation and tail latency before assuming an eight-way partition produces an eightfold economic gain.