AWS Trainium2

AWS Trainium2 is Amazon's third-generation AI training and inference accelerator, and the chip behind Trn2 instances and Trn2 UltraServers. Each chip carries eight NeuronCore-v3 delivering 1,299 FP8 TFLOPS, 667 TFLOPS across BF16, FP16 and TF32, and 181 FP32 TFLOPS, with 96 GiB of device memory at 2.9 TB/s and 224 MiB of software-managed on-chip SBUF scratchpad. Chip-to-chip traffic runs over NeuronLink-v3 at 1.28 TB/s per chip, and 16 dedicated CC-Cores orchestrate collective communication so that reductions do not consume compute cores. AWS publishes a sparse figure of 2,563 TFLOPS spanning FP8, FP16, BF16 and TF32; note that this is 1.97x the dense FP8 rate but 3.84x the dense BF16 rate, so it does not behave like conventional 2:4 structured sparsity and should not be read as a simple doubling. AWS publishes no TDP and does not name the process node or the memory generation for this chip.