Kernel

Every computer has a program that manages all other programs. This program is called the kernel, roughly meaning “core” or “nucleus”. It decides which app is allowed to compute at any given moment, how much memory it gets, and whether it may access the hard drive or the network. A word processor therefore never talks directly to the hardware, but always goes through the kernel. You can picture this like the control center of a train station: the trains don’t set off on their own, but receive clearance from a central point that has an overview of all the tracks. Also important for readers of tech news: the same word is used in a second, related sense with graphics cards. There, a kernel is a small piece of computation that is executed very many times simultaneously.

Why nothing runs without the kernel

The kernel is the only software with full rights over the hardware. All other programs run with fewer privileges. This separation is why a crashing browser doesn’t immediately bring down the entire computer. And it’s why one app can’t simply read out another app’s data.

This is precisely why security vulnerabilities in the kernel are especially serious. Anyone who breaks in there bypasses all of the system’s protective walls. Reports about “kernel exploits” on Android or iPhones always mean: an attacker could take complete control of the device. Manufacturers often release emergency updates for such vulnerabilities within days.

Economically, the Linux kernel is particularly significant. It is freely available and is found in Android phones, in servers at Amazon and Google, and in almost all supercomputers. Thousands of developers from competing companies work on it together. For the AI industry, the second meaning matters just as much: faster GPU kernels directly lower a data center's electricity bill.

From system call to graphics card

A normal program asks the kernel for help via a system call. It says, for example: “Open this file.” The kernel checks the permission, communicates with the storage medium, and returns the data. Such calls happen thousands of times per second without anyone noticing.

Two further tasks are the scheduler and memory management. The scheduler divides computing time into tiny slices and distributes them among the running programs. This creates the impression that a music app, a browser, and an update are all running at the same time. Memory management gives each program its own, isolated area of RAM.

With graphics cards, a kernel works differently. There, it’s a short computational instruction, such as “multiply these two numbers and add the result.” A modern graphics card executes this instruction in parallel across tens of thousands of compute units. This is exactly how the huge tables of numbers that make up AI models are calculated. A common misconception is to treat both meanings as the same thing: the operating system kernel manages, the GPU kernel computes.

Linux, Windows, and the optimizers at Nvidia

In everyday life, the kernel is most often encountered during updates. When Android or a server machine has to restart after an update, the kernel has often been swapped out. The error message “Kernel Panic” on a Mac or Linux machine also belongs here. It means the kernel has found an error it can no longer safely fix.

In tech news, the term also comes up in connection with AI companies. When Nvidia or a startup announces that a model now runs twice as fast, newly written GPU kernels are often behind it. A well-known example is FlashAttention, an optimized kernel for a central computational step in large language models. The models didn’t become smarter as a result, just significantly cheaper to run. Those who can write such kernels are currently among the most sought-after specialists in the industry.