Over 90% of AWS's biggest data-center customers are using its homegrown Graviton chips, an executive says
NEWS | 20 November 2024
An AWS exec says its Graviton CPUs are used by over 90% of its 1,000 largest data-center customers. Graviton chips, launched in 2018, are becoming an important part of AWS's overall business. AWS's custom-silicon strategy includes its Inferentia and Trainium AI chips. In 2018, Amazon Web Services launched Graviton, its line of homegrown central-processing-unit chips for data-center servers. Six years later, the vast majority of AWS's largest server customers have become Graviton users. Rahul Kulkarni, AWS's director of compute and AI/ML, told Business Insider that more than 90% of the 1,000 largest elastic-compute-cloud, or EC2, customers were running Graviton chips. Customers can access Graviton chips only through AWS's EC2 servers. "It just continues to show how Graviton is gaining traction," Kulkarni said. The new data point is the latest indication of Graviton's growing success. AWS previously said that over 50,000 customers, including the 100 largest EC2 users, ran Graviton-based servers. Graviton is a key part of AWS's data-center business. Designing its own chips allows AWS to reduce data-center operating costs because it doesn't have to buy CPUs solely from other vendors, like Intel or AMD. Graviton also uses Arm-based designs, which are more cost- and energy-efficient than conventional x86-powered chips. Amazon doesn't disclose its revenue from its custom-silicon business. The company is scheduled to announce its third-quarter earnings on Thursday. Kulkarni said Graviton's price value, energy efficiency, and general performance drove more customer adoption. Big enterprises such as Epic Games, Databricks, and Pinterest are major Graviton customers, he said. "We have a lot of engagement going on in custom silicon, and that's an area that we will continue to invest in at a very aggressive pace going forward," Kulkarni said. AI inference AWS considered designing custom chips after James Hamilton, a senior vice president and distinguished engineer, wrote an internal six-page strategy document in 2013. It doubled down on this plan by acquiring the Israel-based chip designer Annapurna Labs in 2015. Related stories Bernstein Research said last year that Amazon was the "most successful" designer of Arm-based server chips, supplying over 50% of the chips worldwide. It estimated that Graviton accounted for roughly 20% of AWS's CPU usage as of mid-2022. AWS offers access to other processors, including Intel and AMD chips. Kulkarni said that Graviton was still mostly used for general computing purposes but that a growing number of customers were using it for CPU-based AI inference and machine-learning frameworks. He said that Graviton, now in its fourth generation, had added features and capabilities that enable unique use cases, like AI inference, that don't necessarily need dedicated machine-learning processors. "It's a new revenue channel that goes beyond what AWS had anticipated Graviton to be used for," Kulkarni said. In the past couple of years, Amazon has been cutting costs, shuttering dozens of projects and scaling down unprofitable units. AWS's chip-designing business appears safe from those cutbacks. During an analyst call in August, Amazon's CEO, Andy Jassy, said Graviton had been "very successful" for AWS. He added that AWS's custom AI chips, Inferentia and Trainium, were expected to follow a similar growth trajectory — though they've seen mixed results because of Nvidia's head start. "It's one of the most strategic areas for us," Kulkarni said. "We will absolutely continue to drive innovation in custom silicon as we have been doing for the past 10-plus years." Do you work at Amazon? Got a tip?
Author: Eugene Kim.
Source