NederlandsGPU
A GPU is a chip that performs a huge number of simple calculations at the same time. It was designed for graphics and games, but is now also the most important chip for training and running AI models.
An ordinary processor (CPU) is good at many different tasks, but does them largely one after another. To show an image on a screen, millions of pixels have to be calculated at once. That is what the GPU (graphics processing unit) was designed for: a chip with thousands of small computing cores that together perform the same kind of calculation on very many numbers at once. In a PC it usually sits on a graphics card.
The arithmetic behind AI closely resembles that of graphics: enormous rows of numbers being multiplied together. That is why most AI models run on GPUs, from a single graphics card in a PC to tens of thousands in a data centre. Nvidia is the largest maker. AMD makes them too, and companies such as Google also use their own AI chips.
Why it matters
GPUs are the engine of today's growth in AI: whoever has enough of them can train larger models. They are also useful beyond language models. NVIDIA software uses a single powerful graphics card to work out, in a few milliseconds on average, how a robot arm can move without hitting anything. In a simulated practice world for robots, hundreds of plans run at once on one graphics card. At the same time, some models are becoming so efficient that they no longer need a powerful graphics card and simply run on a phone or laptop.
Articles about GPU
-
Network of artificial neurons learns from its own mistakes
Three researchers describe a learning method for artificial neural networks in Nature: the network works its error backwards and adjusts every connection a tiny bit. As a result, the in-between layers also work out for themselves what matters for the task.
