This document is relevant for: Trn1, Trn1n

Trainium Architecture#

At the heart of the Trn1 instance are 16 x Trainium devices (each Trainium include 2 x NeuronCore-v2). Trainium is the second generation purpose-built Machine Learning accelerator from AWS. The Trainium device architecture is depicted below:


Each Trainium device consists of:

  • Compute:
    • 2x NeuronCore-v2 cores, delivering 380 INT8 TOPS, 190 FP16/BF16/cFP8/TF32 TFLOPS, and 47.5 FP32 TFLOPS.

  • Device Memory:
    • 32GiB of device memory (for storing model state), with 820 GiB/sec of bandwidth.

  • Data movement:
    • 1 TB/sec of DMA bandwidth, with inline memory compression/decompression.

  • NeuronLink:
    • NeuronLink-v2 for device-to-device interconnect enables efficient scale-out training, as well as memory pooling between the different Trainium devices.

  • Programmability:
    • Trainium supports dynamic shapes and control flow, via ISA extensions of NeuronCore-v2. In addition, Trainium also allows for user-programmable rounding mode (Round Nearest Even Stochastic Rounding), and custom-operators via the deeply embedded GPSIMD engines.

This document is relevant for: Trn1, Trn1n