Definition
Plain language
Specialized units on a GPU that do matrix multiplication far faster than the general-purpose math units — but only if your code invokes exactly the right instructions.
As stated in the literature
Dedicated matrix-multiply-accumulate hardware on NVIDIA GPUs; zero tensor-core utilization can mean the kernel doesn't need them, uses an older instruction generation, or silently falls back to scalar code — a distinction only resolvable by reading SASS.
Also called: tensor cores
Why it matters: They offer huge speedups for the matrix math at the heart of AI, but code that fails to invoke them properly leaves that power untapped without any obvious warning.
For example, a matrix-heavy program can run dramatically faster if it triggers these specialized units, but only if the code uses the exact right instructions to engage them.
Heard on the show
“You dumped two hundred numbers on it, most irrelevant to this kernel, and it pattern-matched on the wrong ones — chased a tensor-core "problem" on a kernel that legitimately doesn't need tensor cores.”Episode 177 — Why Raw Profiler Data Made an AI Worse at Writing GPU Code