AWS Trainium3

AWS Trainium3 is Amazon's fourth-generation AI accelerator and, in AWS's own words, "our first 3nm AWS AI chip", made generally available at re:Invent on 2 December 2025. Each chip carries eight NeuronCore-v4 delivering 2,517 TFLOPS of MXFP8 and MXFP4, 671 TFLOPS across BF16, FP16 and TF32, and 183 FP32 TFLOPS, with 144 GiB of HBM3e at 4.9 TB/s and 256 MiB of on-chip SBUF scratchpad. Device-to-device traffic runs over NeuronLink-v4 at 2.56 TB/s per chip, double Trainium2. AWS publishes a sparse figure of 2,517 TFLOPS spanning FP16, BF16 and TF32 but explicitly not FP8; that is 3.75x the dense BF16 rate and exactly equal to the dense MXFP8 rate, so it reflects the chip running sparse work at its low-precision rate rather than any conventional 2:4 doubling. AWS publishes no TDP for the chip.