TMTPOST -- Amazon.com Inc.’s cloud unit Amazon Web Services (AWS) unveiled Trainium3, its first 3-nanometer artificial intelligence (AI) chip, at its annual re:Invent conference Tuesday, intensifying competition with Nvidia Corp. and Google in the lucrative market for AI computing hardware. The cloud computing giant also previewed Trainium4, currently under development, which will support Nvidia's NVLink Fusion interconnect technology for enhanced interoperability.
Related articles
AI Generated Image
The accelerator was recently installed in a few data centers and became available to customers Tuesday, marking a rapid one-year turnaround from its predecessor. "As we get into early next year we'll start to scale out very, very quickly," Dave Brown, an AWS vice president, said in an interview.
Amazon shares rose as much as 2.2% in morning New York trading. Nvidia shares pared gains, while rival Advanced Micro Devices dropped to a session low.
The chip push represents a critical element of Amazon's strategy to stand out in AI. While AWS dominates in rented computing power and data storage, it has struggled to replicate that leadership among AI tool developers, as companies increasingly opt for Microsoft, which maintains close ties to OpenAI, or Google.
Performance Gains Target Cost-Conscious Customers
Trainium3 delivers substantial improvements over its predecessor. Each chip provides 2.52 petaflops of FP8 compute, with memory capacity increased 1.5 times and bandwidth boosted 1.7 times to 144 GB of HBM3e memory. Trn3 UltraServers deliver up to 4.4 times higher performance, 3.9 times greater memory bandwidth and four times better performance per watt compared to Trn2 systems.
The systems scale to 144 Trainium3 chips per server, with capabilities to link thousands of UltraServers providing access to up to 1 million chips—10 times the previous generation. AWS emphasized the chips are 40% more energy efficient than prior versions, promising lower costs for cloud customers.
The upcoming Trainium4 will deliver another significant performance increase and crucially will support Nvidia's NVLink Fusion high-speed interconnect technology, allowing the systems to interoperate with Nvidia graphics processing units (GPUs) while maintaining Amazon's lower-cost server architecture. Amazon provided no timeline for Trainium4's release, though previous patterns suggest details may emerge at next year's conference.
Software Gap Limits Adoption Despite Price Advantage
Amazon's accelerators face a significant hurdle: they lack the deep software libraries that enable quick deployment of Nvidia's graphics processing units. Bedrock Robotics, which uses AWS servers for infrastructure, relies on Nvidia chips for building AI models to guide autonomous construction equipment. "We need it to be performant and easy to use. That's Nvidia," said Chief Technology Officer Kevin Peterson.
Many Trainium chips deployed today serve Anthropic in data centers across Indiana, Mississippi and Pennsylvania. AWS said it had connected more than 500,000 chips for the AI startup and aims to dedicate 1 million chips by year-end. However, Amazon has announced few other major customers, leaving analysts struggling to assess Trainium's market effectiveness. Anthropic also uses Google's Tensor Processing Units and secured a deal earlier this year for tens of billions of dollars worth of Google computing power.
"We've been very pleased with our ability to get the right price performance with Trainium," Brown said, positioning the chips as capable of powering intensive AI calculations more cheaply and efficiently than Nvidia's market-leading processors.
更多精彩内容，关注钛媒体微信号 （ID：taimeiti），或者下载钛媒体 App
