GPU 裏面有甚麼？核心、串流多處理器與記憶體問題

A modern GPU is not one very powerful processor. It is thousands of simple arithmetic cores grouped into streaming multiprocessors, plus specialised tensor cores for matrix maths and ray-tracing cores for light simulation, all surrounded by schedulers, caches and a very wide memory system. The hard engineering problem isn't the maths. It's moving data fast enough to keep all those cores busy.

What are CUDA cores?

Most of the work runs on simple parallel arithmetic units. NVIDIA calls them CUDA cores, AMD calls them stream processors (inside its compute units), and Apple simply counts GPU cores. Each is much simpler than a CPU core, but there are thousands of them running the same instruction across different data at once.

A common confusion: CUDA, ROCm and Metal are software platforms, not hardware. CUDA is NVIDIA's programming model, ROCm is AMD's, and Metal is Apple's graphics and compute API.

What do tensor cores do?

Tensor cores are specialised for matrix multiply-and-accumulate: multiplying two small matrices and adding the result to a third in one step. That one operation is the core of neural networks, so tensor cores are a big part of why GPUs became the foundation for AI training and inference. They also power upscaling features like NVIDIA's DLSS (Deep Learning Super Sampling).

Yes, it's linear algebra all the way down.

What are ray-tracing cores?

RT cores are dedicated hardware for the expensive parts of ray tracing: testing whether a ray of light hits a triangle, and walking the tree structures (bounding volume hierarchies) that make those tests fast. Without them, realistic lighting in real time would be far too slow.

How is a GPU organised?

The structure is hierarchical:

1. Cores do the arithmetic.
2. Cores are grouped into streaming multiprocessors (SMs), each with its own schedulers, registers, shared memory and L1 cache. AMD's equivalent is the compute unit.
3. SMs are grouped into larger processing clusters.
4. All of it sits on a shared L2 cache, memory controllers and the off-chip VRAM.

The schedulers inside each SM switch between groups of threads (NVIDIA calls a group of 32 a warp). When one group stalls waiting for memory, another runs. That's how a GPU hides latency: not by being fast at any single task, but by always having other work ready.

Why does memory bandwidth matter so much?

Doing trillions of operations a second is impressive. Feeding thousands of cores without starving them is harder.

That's the job of GDDR (Graphics Double Data Rate) memory and a very wide memory bus. High-end consumer GPUs now move over 1 TB of data per second between VRAM and the chip. Data-centre GPUs use stacked HBM (High Bandwidth Memory) to reach several TB/s. For many AI workloads, especially LLM inference, memory bandwidth, not compute, is the real limit on how fast tokens come out.

The takeaway

Raw compute is only half the story. The real achievement is coordination: thousands of cores, caches, schedulers and memory channels kept in sync so well that a game frame, a training step or a generated image looks instantaneous.

If you like this kind of under-the-hood breakdown, read how chips got 3,000 times smaller next.

Written by Ali Ahmed · Original article