Amazon has announced the AWS Trainium. It is a chip specially designed for machine learning tasks. The company claims that the chip is faster than competitors in the cloud.
Trainium offers support for TensorFlow, PyTorch and MXNet. According to AWS, instances with Trainium have a 30 percent higher throughput and 45 percent lower cost-per-inference when compared to regular AWS GPU instances.
AWS is also working with Intel to offer EC2 instances based on Habana Gaudi. They should offer a 40 percent better price-performance ratio than the current GPU-based EC2 instances. The chips support TensorFlow and PyTorch, TechCrunch writes.
It is not clear which current GPU offerings Amazon is talking about. Nvidia recently claimed that its A100 GPU achieved the highest scores in the MLPerf benchmark for machine learning. As far as we know, the benchmark has not been run on Trainium hardware yet.
The Trainium instances are available starting next year as EC2 instances and inside Amazon SageMaker. The Habana Gaudi instances will be available in the first half of 2021.