Graviton Processors

When a website or an app runs on the internet, there is a remote computer behind it sitting in a large hall. Such halls full of computers are called data centers, and the largest renter of such computing power is Amazon Web Services, or AWS for short. In the past, AWS bought the computing chips for these computers from other companies, mainly Intel and AMD. Since 2018, the company has been building its own chip: the Graviton processor. It does the same job as a purchased chip, namely processing computing instructions, but is completely tailored to Amazon’s own needs. You can’t buy it; you can only rent it through AWS.

Why Amazon builds its own chips

The main reason is money. A data center consumes enormous amounts of electricity, and a large share of that goes to the processors. Graviton chips need noticeably less energy for the same work than comparable chips from Intel. AWS states for current generations that customers get up to 40 percent better price-performance ratio. Even if that number comes from the marketing department, the direction is undisputed.

The second reason is independence. Anyone who buys all their chips from two manufacturers is dependent on their prices, delivery times, and schedules. A chip of one’s own shifts this balance of power. Amazon can also build in exactly the features its own services need and leave out everything intended for other customers.

For the industry, Graviton is therefore a signal. Google, Microsoft, and Apple are also now designing their own chips. The market for server processors, which Intel dominated for decades, is crumbling on several fronts at once. Analysts estimate that a substantial share of newly installed computing power at AWS already runs on Graviton.

ARM instead of x86: the difference under the hood

Every processor understands a certain language of basic instructions. This language is called an instruction set. Intel and AMD use the x86 instruction set, which has grown since the 1970s and still carries along many old instructions. Graviton instead uses ARM, a significantly leaner instruction set. ARM originally comes from smartphones, where saving power has always been more important than maximum computing power.

Amazon does not design the chips from scratch. The British company ARM sells blueprints for compute cores, which other companies license and adapt. Amazon takes such core blueprints, puts many of them onto a single chip, and adds its own components for memory connectivity and networking. Manufacturing then takes place at a chip factory, in practice mostly at TSMC in Taiwan. So Amazon owns no chip factory, only the design.

The price for this is compatibility. A program compiled for x86 does not run on ARM without further ado. It has to be recompiled, which with modern programming languages is often a matter of pressing a button, but with old specialized software means real work. This exact effort is the most common reason why companies stick with the chips they’re used to.

Graviton in the news and in everyday life

You never see a Graviton processor directly. Indirectly, you use it constantly, since countless websites, streaming services, and apps rent their servers from AWS. Anyone who books a virtual machine there, i.e., a rented computer in the data center, can choose between Intel, AMD, and Graviton. The Graviton variant is usually the cheapest of the three.

In business news, the name usually appears in two contexts. First, in Amazon’s quarterly results, when it comes to the profit margins of the cloud division. Second, in reports about Intel’s weakness, since every Graviton server is a server that Intel did not sell.

A common misconception is that Graviton is an AI chip. It is not. For training large AI models, AWS uses graphics cards from Nvidia or its own specialized chips called Trainium and Inferentia. Graviton does the classic work: databases, web servers, analytics. This unspectacular work, however, is by far the largest part of what happens in a data center every day.