Amazon Unveils Trainium 3 Chips Doubling Performance, Slashing AI Costs by 40%
Amazon Web Services’ Annapurna Labs in Austin is testing its third-generation Trainium AI accelerators using UltraServers with 144 chips, which claim to double Gen2 performance and cut generative AI model development costs by up to 40% versus GPUs. Texas’s low energy costs and tax incentives support AWS’s massive data center expansion.
1. AWS Developing Custom AI Accelerators
Amazon Web Services has designed its own Trainium line of AI-accelerator chips through its Annapurna Labs subsidiary, acquired in 2015. After launching Graviton and Inferentia chips in 2018, AWS debuted the first Trainium in 2020 and has now rolled out Trainium 3.
2. Performance and Cost Advantages
The latest Trainium 3 chips, smaller than a credit card, deliver double the performance of the second generation and can reduce generative AI training and inference costs by up to 40% compared with standard GPUs. UltraServers housing 144 Trainium chips undergo rigorous reliability tests to ensure uninterrupted operation during long AI workloads.
3. Strategic Use of Texas Infrastructure
AWS operates its Trainium-powered data centers in Austin, Texas, leveraging the region’s low energy costs, tax incentives and affordable real estate. These facilities support the heavy computational demands of AI development, enabling AWS to offer high-performance cloud services via its Bedrock platform exclusively to its customers.