Material Insights · Understanding the Material World

Graphics processing unit

Thousands of simple cores work in parallel, first for graphics and now for AI.

Computing and AI · Hardware · No. 291

  • Digital age
  • Readiness 9 of 9
  • Component
Computing and AI · Technology

A chip built to draw video-game scenes turned out to suit the arithmetic that trains modern AI.

Chart: Illustration: Graphics processing unit

Where it comes from

Chart: Map of where it comes from

Graphics chips grew out of the need to draw 3D scenes fast for games. Sony named the GPU in 1994, and Nvidia's GeForce 256 put the whole drawing pipeline on one chip in 1999.

Three things to know
  1. The GeForce 256 of 1999 put transform, lighting and drawing on one chip with 17 million transistors.
  2. Nvidia's GeForce 8800 GTX of 2006 was the first graphics card able to run CUDA programs for general computing.
  3. In 2012 AlexNet was trained on two Nvidia graphics cards in a student's bedroom, a landmark result for neural networks.
The journey
  1. 1 Silicon waferThe chip is made in a leading-edge foundry
  2. 2 Many coresThousands of small cores each work on one pixel or one number
  3. 3 Fast memoryMemory chips sit beside the GPU to feed it data
  4. 4 Card or serverGPUs go on graphics cards for PCs or into AI servers
  5. 5 UseGames, video, science and the training of AI models
Go deeper pick a tab to open it in full