AI High-Performance Chips

Every computer has a main processor that does the computational work. It’s an all-rounder: it can display text, play music, or save files. AI high-performance chips, on the other hand, are specialists. They can only handle a very narrow type of calculation, but they perform it millions of times simultaneously. That’s exactly the kind of calculation artificial intelligence needs — software that learns patterns from examples instead of following fixed rules. This is why large AI systems run almost exclusively on these specialized chips.

Why large language models couldn’t exist without these chips

A language model like ChatGPT consists of billions of adjustable numbers, so-called parameters. For the model to come into being, each of these numbers has to be adjusted many times over. On a normal processor, this would take decades. On thousands of specialized chips, the same work shrinks down to a few weeks. Without this acceleration, the AI wave of recent years simply wouldn’t have happened.

That’s why these chips have become an economic and political power factor. The US manufacturer Nvidia dominates the market and, as a result, briefly became the most valuable company in the world. Individual server cards cost several tens of thousands of euros, depending on the model. Anyone wanting to train a large model needs tens of thousands of them.

Because these chips are so important, the US has restricted their export to China since 2022. Power consumption is also an issue. A large data center full of AI chips uses as much energy as a small city. Debates about electricity prices and new power plants are now directly tied to this technology.

Computing with many small workers instead of a few smart ones

Surprisingly little variety happens inside an AI model. Almost everything boils down to multiplying and adding large tables of numbers, so-called matrix multiplications. A main processor might have 16 highly flexible computing cores. An AI chip, by contrast, has tens of thousands of very simple computing units, all working in parallel on the same type of task.

A comparison helps: a main processor is like a professor who can solve any task at all, but only one at a time. An AI chip is like a hall filled with a thousand students, all simultaneously calculating simple multiplications. For tax returns, the professor is better. For ten million multiplications, the hall wins.

An often underestimated point is memory. The computing units are usually faster than data can be delivered to them. That’s why modern AI chips have a special fast memory sitting right next to the computing core. These components are called HBM, short for High Bandwidth Memory. They are currently more often the bottleneck than raw computing power itself.

From the graphics card in a PC to the chip in a smartphone

The best-known representative is the graphics card, or GPU for short. It was originally built for video games, because they too require many simple calculations at once. That it turned out to be well-suited for AI was a lucky coincidence. Today, Nvidia sells server versions of these cards that no longer even have a display connector.

Alongside these, there are chips designed from the ground up purely for AI. Google builds its TPUs, while Amazon and Microsoft develop their own variants for their data centers. A small AI chip is even built into smartphones. It ensures that facial recognition or voice commands work directly on the device, without an internet connection.

These chips show up in economic news almost daily. Reports on supply shortages, export bans, or Nvidia’s quarterly earnings move entire stock indices. A common misconception here is that a better chip automatically means a smarter AI. The chip only delivers computing power. What the model ultimately can do depends just as much on the training data and the model’s architecture.