Amazon has announced the upcoming release of its next-generation AI training chip, Trainium3, which is set to debut by late 2025. This chip marks a significant advancement as it will be the first of Amazon's self-developed chips to utilize TSMC's cutting-edge 3nm fabrication process. According to a report from Economic Daily News, Trainium3 will enhance performance and efficiency compared to its predecessor.
During the recent AWS Cloud Conference, CEO Matt Garman highlighted that Trainium3 is expected to be twice as fast as Trainium2 while boasting improved energy efficiency by 40%. In addition to this performance boost, Amazon's UltraServers, powered by Trainium3, are anticipated to provide four times the performance of those using Trainium2 chips.
Amazon also unveiled new data center servers featuring the Trainium2 chips, which have attracted notable clients such as Apple, confirming its adoption of the technology for machine learning and AI applications. The new servers are part of a competitive strategy, aiming to rival NVIDIA’s flagship systems, which are equipped with 72 advanced Blackwell chips.
Gadi Hutt, AWS’s head of AI chip business development, pointed out that AWS has the capability to connect a larger number of chips than NVIDIA, offering potential advantages in computational power. He noted that Trainium2 could deliver greater processing capabilities while also reducing costs, with some AI models being trained at up to 40% lower expenses compared to using NVIDIA's chips.
Both Amazon and NVIDIA depend on TSMC, Taiwan's leading foundry, for their chip production needs, indicating the critical role TSMC plays in the semiconductor supply chain. With Trainium3, Amazon aims to strengthen its position in the AI training market, demonstrating its commitment to enhancing computing efficiency and performance through innovative chip technology.
