Amazon's cloud division, Amazon Web Services (AWS), unveiled its latest AI chip, Trainium2, and a new general-purpose processor, Graviton4, at the Reinvent conference in Las Vegas. In addition to these, AWS will provide access to Nvidia's latest H200 AI graphics processing units. AWS aims to offer a diverse range of cost-effective cloud solutions, not limited to its own products but also including premium offerings like Nvidia's GPUs.
This strategy mirrors Microsoft's recent move in the cloud sector, where it introduced its first AI chip, Maia 100, and also plans to incorporate Nvidia H200 GPUs in its Azure cloud service.
Graviton4, based on Arm architecture, is more energy-efficient than its Intel or AMD counterparts and offers 30% better performance than Graviton3. This enhancement comes at a time when inflation and rising interest rates are motivating organizations to seek more cost-effective cloud solutions.
AWS's Graviton chips are already in use by over 50,000 customers. Companies like Databricks and Anthropic are planning to utilize the new Trainium2 chips for model building. These chips promise a fourfold improvement in performance over their predecessors.
AWS also announced its operation of over 16,000 Nvidia GH200 Grace Hopper Superchips for Nvidia’s R&D, though these won't be available to other AWS customers. The demand for Nvidia GPUs has surged, particularly since the release of OpenAI's ChatGPT, leading to a shortage of these chips.
While AWS's introduction of an AI chip might seem competitive to Nvidia, Amazon is also expanding its collaboration with the chipmaker. This move offers AWS customers an additional AI computing option, especially if they can't secure the latest Nvidia GPUs.
Amazon has been a leader in cloud computing for over a decade, with a history of offering GPU rentals in its cloud. Following competitors like Alibaba and Google, AWS released its own in-house developed AI processor in 2018, providing powerful computing at an affordable price.
Since launching its EC2 and S3 services in 2006, AWS has introduced over 200 cloud products. While not all have been successful, Amazon continues to invest in its Graviton and Trainium programs, indicating a strong market demand.
AWS hasn't specified release dates for virtual-machine instances using Nvidia H200 chips or those based on Trainium2, but customers can start testing Graviton4 virtual-machine instances, which are expected to be commercially available in the coming months.