An Easy Introduction to CUDA

Most people who hear 'GPU programming' assume it's for machine learning researchers or graphics engineers. It's not. Any time you're running the same operation on a large array of data, a GPU can do that work orders of magnitude faster than a CPU. The gap isn't marginal. In this course, you go from a single-threaded kernel that takes 75 milliseconds to a prefetched multi-block kernel that finishes in under 50 microseconds on an NVIDIA T4. Same math, same hardware, completely different result. The course builds in a straight line. First you learn why CPUs and GPUs are built differently and what that means for the kinds of problems each one handles well. Then you learn how CUDA threads and blocks work, how to give every thread its own slice of the array, and why block sizes should be multiples of 32. By the end of the second unit you'll have a working parallel kernel. Then the third unit shows you the part most tutorials skip: why that parallel kernel is still slow, how Unified Memory page faults are secretly copying data behind your back, and how one function call fixes it. You don't need a background in computer architecture or prior GPU experience. You need to be comfortable writing C-style code and curious about why programs run at the speed they do. Every concept here is tied to a concrete number you can measure yourself, so you're not taking anyone's word for what's fast.