수학 연산에서 반복문이 느린 이유

Coming from plain Python, you reach for a for loop to add two lists number by number. It works, but for math at scale it is the slow path. 🐢

Every loop step in Python carries interpreter overhead: type checks, object boxing, and bytecode dispatch happen again and again for each element.

A neural net touches millions of numbers per pass. Multiply that tiny per-element cost by millions and your loop crawls while real work stalls.

Vectorization means describing the whole operation at once on a tensor, so the heavy lifting drops into fast compiled C and CUDA code under the hood.

This loop adds two tensors element by element in Python. Correct, but it pays the interpreter tax on every single step.

The same result in one vectorized line. PyTorch loops in compiled code, not in the slow Python interpreter.

That single expression hands the whole array to an optimized kernel. It runs the loop for you, far closer to the hardware and far faster.

Tensors store numbers in one tight, contiguous block of memory. The CPU streams them in cache-friendly order, something a Python list cannot promise.

Modern chips use SIMD instructions that apply one operation to several numbers in a single clock tick. Vectorized code unlocks this; loops usually do not.

A GPU has thousands of cores hungry for parallel work. Feed it whole tensors and it shines; feed it one element at a time and it sits mostly idle.

The mindset shift: stop asking what happens to one number and ask what happens to the whole array. That question is the key to fast deep learning code.

Python loops pay a per-element tax that vectorized tensor ops avoid by running in compiled, SIMD-ready, GPU-friendly code. Think in arrays, not single numbers. ✅