Grace CPU

Every computer has a main processor, that is, the chip that carries out the general computing steps and controls the rest of the system. Grace CPU is one such main processor made by the company Nvidia. For decades, Nvidia was known mainly for graphics chips, meaning components that handle a great many simple calculations simultaneously. With Grace, the company has also been building the chip alongside it itself since 2023. Grace is not found in laptops or game consoles, but in large data centers, that is, in halls full of servers. The name is a reference to Grace Hopper, an American computer scientist and pioneer of programming languages.

Why Nvidia is suddenly building its own main processors

In an AI server, two types of chips work together. The graphics chips take on the actual computational work for the AI model. The main processor loads data, distributes tasks, and keeps the operating system running. If this supplier is too slow, the expensive graphics chips have to wait. That was precisely a bottleneck for years, and Grace is meant to eliminate it.

There is also an economic motive behind this. Until now, the main processors in such servers came almost always from Intel or AMD. With Grace, Nvidia now sells a complete package and earns money on the whole server instead of just one part. For investors, this is an important point: the share Nvidia earns per server sold rises significantly as a result. At the same time, the classic processor manufacturers lose part of a growing market.

A second reason is power consumption. Large data centers are running up against the limits of their power connections. Grace is designed for as much computing performance per watt as possible. Anyone who can compute more for the same electricity bill saves a great deal of money in operation.

Design: ARM cores and a chip duo

Grace does not use the architecture of classic PC processors, but that of ARM. ARM is a British company that develops processor blueprints and licenses them to other manufacturers. This design is also found in nearly every smartphone and is considered especially power-efficient. Nvidia buys the license and designs from it its own chip with 72 computing cores. A computing core is an independent computing unit; many cores can work in parallel.

What’s special is the connection to the outside. Grace is usually packaged tightly together with a graphics chip into a shared component. These combinations are called Grace Hopper or Grace Blackwell, depending on the generation of the graphics chip. Between the two lies a very wide data link called NVLink. It transfers data many times faster than the usual plug connections on a motherboard.

This allows both chips to access the same working memory without constantly sending copies back and forth. That is the actual trick. A language model with hundreds of billions of parameters otherwise barely fits into the memory of a single graphics chip. A common misconception is that Grace is a competitor to Nvidia’s graphics chips. It is their supplier and partner, not a replacement.

Grace in supercomputers and in the quarterly figures

You won’t buy a Grace chip directly. You use it indirectly as soon as you ask an AI chatbot a question or have an image generated. Some of these requests run on servers with Grace chips. Several European supercomputers for weather and climate research also rely on it, such as the Alps computer in Switzerland.

In business news, the name usually appears as part of a package. When GB200 systems are mentioned, the G stands for Grace, the B for the Blackwell graphics chip. Such systems cost millions per server rack, and delivery volumes are seen as a gauge for the entire AI boom.

Grace is also an example of a trend: major tech companies are increasingly developing their most important chips themselves. Amazon, Google, and Apple do the same. Whoever controls the chip is less dependent on suppliers and can fine-tune hardware and software precisely to match each other.