Explainer
CPUs and GPUs are designed around different kinds of work. A CPU handles a broad mix of tasks, while a GPU can apply similar operations to many pieces of data at once. That makes GPUs useful for many deep-learning calculations, including matrix multiplication, when the workload is large enough to benefit.
The key ideas
- CPUs handle varied control and computation tasks.
- GPUs can process many similar operations in parallel.
- Memory movement and arithmetic are different costs.
Explore the subject
Original recording: TikTok, 2020-10-17.
Video transcript
Read the original recording’s transcript
Automatically transcribed from the original recording. Small transcription errors may remain.
Why are GPUs better than CPUs for artificial intelligence and deep learning? Well, in a nutshell, CPUs like the Intel or AMD processor and the device you are using to watch this video right now is best at handling single, more complex calculations sequentially. GPUs like the ones produced by NVIDIA for video game graphics and 3D software are better at handling multiple, more straightforward calculations in parallel. So in other words, the CPU excels at fetching small amounts of memory or data quickly, while a GPU is good at bringing large amounts of
memory, something we call matrix multiplication. So a CPU could be represented by imagining a sports car that can quickly move from point A to point B, but can only carry a tiny load, therefore it will have to make multiple trips. Now imagine our GPU as a huge diesel truck, not quite as fast, but can take much larger loads in one trip versus the car.
