Amazon Unveils AI chip, Enters Market with Microsoft and Nvidia

At the re: Invent conference in Las Vegas, Amazon revealed Trainium2, an artificial intelligence (AI) processor designed for training big language models. This development coincides with the debut of processors by other tech giants and chip manufacturers, including as Microsoft, Nvidia, and AWS.

How good is Amazon AI chip?

Amazon claimed that the Trainium2 chip can provide up to 4 times faster training than the first generation Trainium chips and it offers up to 2 times the energy efficiency. Enterprises can deploy these chips in Amazon Elastic Compute Cloud (Amazon EC2) UltraClusters (up to 100,000 chips).

"Trn2 instances are designed to enable customers to scale up to 100,000 Trainium2 chips on next-generation EC2 UltraClusters connected to the AWS Elastic Fabric Adapter (EFA) petabit network, delivering up to 65 exaflops of compute and giving customers access to supercomputer performance. When necessary," the company said. 

Amazon will start offering Trainium2 chips next year. The company continues to offer access to Nvidia's latest H200 AI graphics units. 

Amazon releases AWS Graviton4 

Amazon also announced the Graviton 4 chip, the company's number the most powerful and energy-efficient AWS processor to date for a variety of cloud workloads. According to the company, Graviton4 offers up to 30% better computing performance, 50% more cores and 75% more memory bandwidth than the current generation Graviton3 processors.

Microsoft and Nvidia latest chips

The development comes a few weeks after Microsoft announced the Microsoft Azure Maia 100 AI chip at Ignite 2023. The chip is optimized to handle AI tasks and creative artificial intelligence. It also announced the Microsoft Azure Cobalt CPU, an Arm-based processor that runs general-purpose computing workloads in the Microsoft cloud. Nvidia, on the other hand, offers H100 GPUs that are selling like hotcakes.

The company has been developing chips for the Chinese market even after the US tightened chip export rules. Meanwhile, AMD is also struggling with its Instinct MI300X processor, which it says can handle multi-GPU work.