Amazon Web Services is planning an update to its Graviton4 chip with 600 gigabits per second of network bandwidth. The cloud provider is positioning the chip as an alternative to Nvidia’s expensive GPUs in the AI landscape.
That’s according to CNBC. The biggest challenge for AWS is in AI infrastructure, where Amazon has to compete with Nvidia’s dominance. At the AWS re:Invent 2024 conference in December, the company announced Project Rainier, an AI supercomputer for Anthropic, in which AWS is investing $8 billion.
Gadi Hutt, Senior Director for Customer and Product Engineering at AWS, emphasizes that Amazon wants to reduce AI training costs. Nvidia’s GPUs are expensive, prompting companies to look for alternatives. Anthropic’s Claude Opus 4 AI model runs on AWS’s Trainium2 chips. Project Rainier is powered by more than half a million of these chips.
The Graviton4 chip is central to Amazon’s custom chip strategy. Ali Saidi, distinguished engineer at AWS, compares the speed of 600 gigabits per second to a machine that can read 100 music CDs per second.
Amazon’s entire chip portfolio comes from Annapurna Labs in Austin, Texas. In addition to Graviton4, this facility also develops other chips for AWS. Earlier this year, AWS introduced new EC2 instances based on Graviton4, which significantly improves performance compared to the previous generation.
Next generation already on the way
AWS is already looking ahead to the next step. Trainium3 will be released this year and promises double the performance of Trainium2. At the same time, energy consumption should be reduced by 50 percent. Demand for these chips already exceeds supply, according to Rami Sinno, director of engineering at Annapurna Labs.
For Amazon, this is part of a broader strategy to control the entire AI infrastructure stack. The company aims to have complete control over everything, from networking to training and inference. The Graviton4 update will be available before the end of June, according to an AWS spokesperson.