AWS Trainium is Amazon's custom AI chip, built for training and running AI models. The current generation, Trainium3, has 144 GB of HBM3e memory with 4.9 TB/s of bandwidth per chip.
What AWS publishes
AWS's Trainium page calls it a "purpose-built AI chip" for training and inference at scale. Trainium3 is described as having up to 2x MXFP8 compute throughput over Trainium2 and 1.7x its memory bandwidth, and UltraServers scale up to 144 Trainium chips. The Trn3 instance page gives 144 GB of HBM3e per chip and up to 20.7 TB of HBM3e per UltraServer.
Software is the Neuron SDK, with the Neuron Kernel Interface (NKI) for low-level kernels plus profiling and monitoring tools. See what the Neuron SDK is.
Where it fits
Trainium targets both training and inference, which separates it from Inferentia. The head-to-head with NVIDIA is in AWS Trainium vs NVIDIA, and the wider map is NVIDIA's main competitors in AI.
