A modern GPU is not one big processor. It is a large number of identical, smaller ones working side by side. In Nvidia's design, each of those smaller units is called a streaming multiprocessor, or SM.
What an SM does
According to Nvidia's CUDA programming guide, a GPU can be treated as a collection of SMs, organized into groups called graphics processing clusters. When a program runs a task on the GPU, it launches many threads, often millions, and groups them into blocks. All the threads of one block run on a single SM, and blocks are handed out to whichever SMs are free, in no guaranteed order.
Picture a large kitchen with many identical stations. A big order is split into batches, and each batch goes to one station. More stations mean more batches cooked at once.
Each SM contains its own set of arithmetic units, a small pool of fast local memory and the scheduling logic that keeps them busy. For AI work, the SMs are also where the matrix math at the heart of a neural network is carried out.
Why the count matters
The number of SMs differs between GPU models and generations. A chip with more SMs can run more blocks at once, so it generally does more work per second, all else equal. It also changes how software splits up a job. Libraries often choose their work layout based on how many SMs the card has, so the same program can behave slightly differently on different chips.
That side effect is why the count can show up in places you would not expect. For example, how some numerical routines divide up their work, and therefore the exact order of floating point operations, can depend on the GPU model.
What it does not tell you
SM count alone does not rank two chips. Memory capacity and speed, covered under high-bandwidth memory, the links between chips, clock speeds and the design of each generation all matter. For inference, memory bandwidth is often the limiting factor, not raw compute.
How it relates to other terms
An SM is one piece of what people loosely call a GPU. The programming model that sends work to it is CUDA. Other vendors have equivalent units under different names, so the idea carries over even when the label does not.
If you are comparing hardware, treat the SM count as a clue about the chip's size, then look at memory and interconnect before drawing a conclusion.