NVLink-C2C

A computer contains multiple computing components that constantly need to exchange data. Between them lie data links, and these are often the slowest part of the system. NVLink-C2C is one such link, developed by the company Nvidia, the largest manufacturer of graphics and AI chips. The abbreviation C2C stands for “Chip to Chip.” The connection couples two components that sit closely side by side within the same housing, at extremely high speed. Up to 900 gigabytes per second flow across it, roughly equivalent to about 200 feature films in HD quality per second.

Why the path between two chips becomes a bottleneck

A large AI model consists of many billions of stored numbers, so-called parameters. These numbers must be transported from memory to the processing unit for every calculation. If the model doesn’t fit into the memory of a single chip, it must constantly be shuffled back and forth between components. If the connection between them is too narrow, the expensive computing chip spends most of its time simply waiting for data.

This is exactly where NVLink-C2C comes in. The classic connection in computers is called PCI Express, or PCIe for short, which in current versions achieves about 64 gigabytes per second. NVLink-C2C is thus roughly fourteen times faster. This difference determines whether an AI model can run meaningfully on a system at all.

The second advantage is economic in nature. Data centers pay for electricity and for the time their chips spend working. A chip that is waiting for data consumes power without accomplishing anything. Faster connections therefore increase utilization and lower the cost per computing task.

Short distances, shared memory

The most important trick is distance. Normal connections run across the computer’s main circuit board and cover several centimeters. NVLink-C2C connects two chips that sit only millimeters apart on the same substrate. Short distances mean fewer disturbances, lower power consumption, and more possible lines running side by side. Nvidia states that the connection requires about five times less energy per transmitted data packet than PCIe.

On top of that, there’s a special feature in the software. Both chips see their working memory as a single, contiguous area. This is called coherent memory: if one chip changes a value, the other notices the change immediately. Programmers therefore no longer need to copy data by hand — they simply access it.

A comparison makes this clear. Otherwise, two separate chips work like two colleagues in different offices who send each other folders by courier. With NVLink-C2C, they sit at the same desk and look at the same documents. Incidentally, a common misconception is confusing NVLink-C2C with regular NVLink. Regular NVLink connects entire graphics cards via cables within a server rack, whereas C2C operates within a single component.

Grace Hopper and the chips in the quarterly earnings

NVLink-C2C is best known from Nvidia’s superchips. In the Grace Hopper model, the technology connects a main processor called Grace with an AI computing chip called Hopper. In the successor, Grace Blackwell, it additionally couples two computing chips into a single unit. Such components are found in the servers used to train and operate large language models.

In business news, the term usually appears indirectly. When reports discuss Nvidia’s record revenues or supply bottlenecks for AI servers, it is precisely these products that are being referred to. NVLink-C2C is one of the reasons why competitors find it hard to catch up: the connection is not an open standard but belongs to Nvidia. Anyone who wants to use it has to buy from Nvidia.

You won’t find this technology in your own laptop or phone. It only pays off for very large computing tasks. Nevertheless, it touches everyday life: every chatbot response is generated on servers in which such connections determine the flow of data.