A Machine Learning Specialist has trained an Apache MXNet model using Amazon SageMaker. The Specialist wants to accelerate his inference workloads without having to pay for expensive GPU-based instances. Which is the MOST cost-effective solution for this problem?