The Arm chip architecture took the mobile world by storm in 2007 after the release of the first iPhone.
Just two years later, an Amazon executive who now leads the company’s semiconductor development, believed Arm would eventually be a big part of server-side computing.
“I’ve observed, over the years what happens in mobile ends up happening in servers. It just takes five to 10 years,” said James Hamilton, senior vice president and distinguished engineer at Amazon, during a talk at Amazon’s Silicon Innovation Day earlier this month.
“I blogged this back in 2009. And it’s not a complicated story,” Hamilton said.
That was a springboard for Amazon to start thinking about its first Arm chip, which Hamilton called a dongle, which was more a server-attached chip on which security, networking and virtualization tasks could be offloaded. A goal with the chip was to free up server resources and sell more cores on AWS, Hamilton said.
Hamilton wanted a custom Arm processor and ultimately worked with Cavium to deliver its first custom chip called Nitro in 2013. The chip is now the highest volume semiconductor part at AWS, with over 20 million installed since it was introduced.
Hamilton saw the value of silicon customization as AWS moved to bare-metal offerings. In 2013, he shared his ideas on the need to homegrown Amazon chips to the absolute top of the company.
“So back in 2013, I wrote a doc called AWS custom hardware that was reviewed with Andy Jassy and Jeff Bezos on this twin thesis: thesis one … Arm volume will yield a great server processor. Thesis two: server innovation is going to all be on a chip. Conclusion: AWS needs to do a custom processor,” Hamilton said.
Soon after, Amazon got in touch with chipmaker Annapurna Labs for Nitro v2, the successor to the original Nitro. The companies were able to work quickly to launch the product, and in January 2015, Amazon acquired Annapurna Labs.
Amazon kept its faith in Arm-based server processors despite some top Arm server chipmakers, including Cavium and Calxeda, dropping out of the race. After Nitro, Amazon expanded its Arm silicon development into other lines that include Graviton, which was announced at Reinvent 2018 and was delivered to market as part of its scale-out workload offerings; Inferentia, for high-performance inference in the cloud, Trainium, which was introduced in 2020.
Amazon has backed Arm since its entry into the semiconductor space because of the volume shipments, which has strengthened the ecosystem and software development. Arm-based chip shipments totaled 7.4 billion in the most recent quarter, growing by 7 percent compared to the same quarter last year.
“This industry is all about volume. Anytime you can do something at scale, it opens up more R&D investment. You can do more for customers,” Hamilton said, adding “Arm’s mobile and IoT volumes are going to feed the R&D investment necessary to produce great server processors.”
AWS also hosts servers with x86 chips from Intel and AMD. Research firm Trendforce in a study said Arm architecture’s market share in datacenters could reach 22 percent by 2025.
Arm server cores are designed for cloud-native applications that can scale up and served quickly. Cloud providers such as Google and Oracle have opted for Arm-based servers from Ampere, whose chips called Altra have up to 128 cores.