Cloud Giants Intensify Custom Silicon Race
Amazon Web Services has formally taken the wraps off Trainium3, its third-generation custom silicon accelerator tailored for large-scale generative model training and inferencing workloads.
Built on a 3-nanometer process node by TSMC, Trainium3 introduces direct-to-chip liquid cooling architectures that enable 64-chip server enclosures with 4x the FLOPS density of previous AWS generations.
NeuronLink Interconnect & Anthropic Partnership
Using the proprietary NeuronLink interconnect fabric, AWS can cluster up to 100,000 Trainium3 chips into unified UltraClusters capable of delivering exaflops of AI compute. Anthropic has already committed to training its future foundational models on the new architecture.