Inferentia is AWS's chip line for inference. Trainium is its newer line, which AWS positions for both training and inference.
Side by side, from AWS
| Inferentia2 | Trainium3 | |
|---|---|---|
| Positioned for | Inference | Training and inference |
| Memory per chip | 32 GB HBM | 144 GB HBM3e |
Sources: the Inferentia page (two second-generation NeuronCores and 32 GB of HBM per Inferentia2 chip; the first generation had 8 GB of DDR4) and the Trainium page (144 GB of HBM3e at 4.9 TB/s).
One software stack
Both run on the Neuron SDK. AWS says the SDK trains models on Trainium and deploys them on Inferentia, and that it integrates natively with frameworks such as PyTorch and TensorFlow. More in what the Neuron SDK is.
For the Trainium chip on its own, see what AWS Trainium is. The rest of the field is in NVIDIA's main competitors in AI.
