What is a GPU Core?

A GPU core is one of the compute units that make up a Streaming Multiprocessor. The two kinds that matter for AI workloads are CUDA Cores, which handle general scalar arithmetic, and Tensor Cores, which are wired specifically for matrix multiply-accumulate.

Reading "core" the way you'd read a CPU spec sheet is misleading. A CPU core fetches its own instructions and can run an entirely different program from its neighbor. A GPU core is closer to a fixed-function pipe: data goes in, one kind of operation happens (a floating-point multiply-add, say), and a result comes out, on a schedule set elsewhere.

That "elsewhere" is the Streaming Multiprocessor. It owns the register file, the instruction scheduler, and the pool of cores underneath it, so the SM, not the core, is what actually plays the CPU-core role: holding state and issuing instructions. A core's identity comes from what instruction it executes and how fast, which is why CUDA Cores and Tensor Cores are worth telling apart rather than folding into one "core count" on a spec sheet.

Two GPUs with the same total core count can differ a lot in real throughput depending on that split; see the Aquanode marketplace for the CUDA Core and Tensor Core counts behind each listed model.