AWS Trainium & Inferentia
AWS-designed silicon for cheaper training and inference.
❓ What is it?
Purpose-built machine-learning chips: Trainium (Trn instances) accelerates model TRAINING; Inferentia (Inf instances) accelerates INFERENCE — both targeting better price-performance than general GPUs, programmed via the AWS Neuron SDK.
💡 Why does it exist?
At scale, compute is the dominant AI cost. Custom silicon tuned for tensor workloads delivers the same work for less money and energy — AWS's answer to GPU supply and cost pressure.
⏱️ When should you use it?
Consider them when training or serving large models with cost as a driver and your framework stack (PyTorch/TensorFlow via Neuron) is compatible; stay on GPUs when a dependency demands CUDA.
🗺️ Where does it fit?
As instance choices inside SageMaker and EC2; Bedrock also runs partly on this silicon behind the scenes, which is one reason its token pricing keeps falling.
🔌 How do you integrate it?
Pick trn1/inf2 instance types, compile the model with the Neuron SDK, and benchmark against the GPU baseline before committing fleets.
🧩 Commonly integrated with
🎯 Exam angle (AIF-C01)
- Names encode purpose: TRAINium trains, INFERentia infers. Free marks if you remember that.
- They are the answer to "reduce cost / improve energy efficiency of large-scale training or inference on AWS".