AWS Unveils Trainium Chip to Accelerate and Cut Machine Learning Costs
AWS announces Trainium, a custom silicon chip designed specifically for training machine learning models in the cloud. The new hardware promises significantly faster speeds and lower costs compared to current GPU instances.
AWS announces Trainium, a custom-designed chip dedicated to training machine learning models in the cloud. The company reveals that this next-generation hardware delivers 30% higher throughput and reduces cost-per-inference by 45% compared to standard GPU instances. Developers access this power through new EC2 instances and directly within the Amazon SageMaker platform.
In addition to Trainium, AWS partners with Intel to launch EC2 instances powered by Habana Gaudi chips. These Intel-based instances offer up to 40% better price performance than current GPU options for machine learning training. Both new chip solutions support popular frameworks like TensorFlow and PyTorch.
These new training chips complement AWS Inferentia, the custom chip the company launched last year for machine learning inferencing. Because Trainium uses the same software development kit as Inferentia, developers easily create an end-to-end machine learning pipeline. AWS plans to make both new training instance types available in the cloud during the first half of 2021.