Overview
- Trainium3 is available to customers now and already installed in select data centers, with AWS planning a rapid capacity ramp early next year, according to vice president Dave Brown.
- AWS claims more than four times the compute of the prior generation with about 40% lower energy use and up to 50% lower training and operating costs versus GPU-based systems.
- New UltraServer systems pack 144 Trainium3 chips and can be linked by the thousands to present up to one million chips to a single application, using 3 nm silicon with expanded HBM3e memory and bandwidth.
- AWS disclosed Trainium4 is in development with support for Nvidia's NVLink Fusion interconnect and a stated goal of at least triple Trainium3 performance, with no release timeline announced.
- Anthropic is a flagship user with more than 500,000 Trainium chips interconnected and a plan for one million by year-end, while many buyers still favor Nvidia's software ecosystem and Nvidia retains an estimated 80–90% market share.