TheTechno Optimist.

StoryHardware

AWS launches Trainium3 UltraServers: first 3nm AI accelerator, 144 chips per server, 362 PFLOPS FP8, 50% cost savings vs. GPU

AmazonTrainium (Trainium3)

AWS announced general availability of Trainium3 UltraServers at re:Invent in early December 2025. Each Trainium3 chip delivers 2.52 PFLOPS of FP8 compute on 3nm silicon with 144 GB of HBM3e memory. A single UltraServer with 144 chips reaches 362 PFLOPS FP8, reducing training costs by up to 50% versus GPU-based systems while improving performance by up to 4.4x.

  • Trainium3 is AWS's first AI chip built on 3nm process; 2.52 PFLOPS FP8 per chip.
  • UltraServer density: 144 Trainium3 chips per system, delivering 362 PFLOPS FP8.
  • 1.5x more memory, 1.7x more bandwidth, and 5x larger HBM3e capacity (144 GB) versus Trainium2.
  • Cost improvements up to 50%, performance up to 4.4x higher than equivalent GPU systems for AI training.
Read the original · Press Computing
Primer