Amazon launches Trainium3 AI accelerator, competing directly against Blackwell Ultra in FP8 performance — new Trn3 Gen2 UltraServer takes vertical scaling notes from Nvidia's playbook

Amazon Web Services has unveiled its Trainium3 AI accelerator, boasting twice the speed and four times the efficiency of its predecessor. The Trainium3 processor offers up to 2,517 MXFP8 TFLOPS, competing directly with Nvidia's Blackwell Ultra. The Trn3 UltraServer, equipped with 144 Trainium3 chips per rack, matches Nvidia's NVL72 GB300 in FP8 performance. AWS's Trainium3 features dual-chiplet architecture with 144 GB of HBM3E memory and NeuronCore-v4 cores, providing peak memory bandwidth of 4.9 TB/s. Additionally, AWS introduced updates to its Neuron software stack to enhance developer accessibility and performance control on Trainium platforms.