CPU กับ GPU กับ MPS: เลือกอุปกรณ์

Every tensor lives on a device: the CPU, a GPU, or Apple's MPS. The device decides how fast your model trains. ⚡

Your CPU is the safe default. It runs every PyTorch operation, just more slowly on the heavy matrix math that big networks demand.

A GPU does thousands of multiplications in parallel, making it far faster for deep learning. PyTorch reaches NVIDIA GPUs through CUDA.

On a Mac with Apple Silicon, MPS taps the built-in GPU. It is the fast path for M-series chips without any NVIDIA hardware.

Ask PyTorch whether an NVIDIA GPU is available before you use one. cuda.is_available returns True or False so you never assume.

On a Mac, check the Apple backend the same way. mps.is_available tells you if the M-series GPU is ready to use.

A clean pattern tries CUDA first, then MPS, then falls back to CPU. This device string adapts to whatever machine runs your code.

Send data to the chosen device with .to(device). The tensor's numbers are unchanged, but its math now runs there.

Your model and your data must share one device. Mixing CPU and GPU tensors throws a device mismatch error during the forward pass.

Not sure where data sits? Read the .device attribute. It prints cpu, cuda, or mps so you can confirm placement at a glance.

Use the GPU for real training, and the CPU for quick tests or tiny data. The fastest device is the one your hardware actually has.

Pick the right call for a Mac with Apple Silicon.

You learned to detect CUDA, MPS, or CPU, pick the best one, and move tensors there. Same code, fastest hardware available. 🚀