Amazon Web Services (AWS) gets another innovation besides the recently introduced macOS instances based on Mac minis. It is about the Trainium, a chip for the cloud computing division of Amazon that was developed explicitly for training neural networks. Like the Inferentia ASICs and the Graviton2 CPUs, the Trainium is an in-house development.
Customers should be able to use Trainium from the second half of 2021. According to Amazon, the focus in development was on a very good performance in relation to price. According to Amazon, the instances should be below those with Intel's Habana Gaudi, which in turn already have a 40 percent better price-performance ratio than the p4d.24xlarge configuration with eight Nvidia A100.
Technical details unknown
Amazon has not yet published details about the technology of the Trainium chip and its performance. The only known fact is that the AWS instances will have the highest teraflops performance of all the company's machine learning instances.
It is also known that the Cloud Standard frameworks that support Pytorch, MXNet and Tensorflow. The software development kit (SDK) called Neuron is also already known from the Inferentia ASICs.
Trainium is available in the Sagemaker service and in the EC2 instances (Elastic Compute Cloud). Learning algorithms and frameworks frequently used in AWS should be pre-installed.