Tensor Core
A dedicated unit that computes a small matrix multiply-accumulate every clock — the engine behind the headline "AI TFLOPS".
Deep learning is mostly matrix math, so these cores — running at lower precisions such as FP16, FP8 and FP4 — deliver the headline “AI TFLOPS”. NVIDIA’s Transformer Engine dynamically picks precision to keep them saturated.
A Tensor Core performs a fused matrix multiply-accumulate (D = A×B + C) on small tiles every cycle. Stacking thousands of them is what turns a GPU into an AI engine: a transformer’s forward and backward passes are overwhelmingly dense matmuls.
In the overview matmul, each Streaming Multiprocessor feeds blocks of A and B into its Tensor Cores as it marches along the shared dimension.
TensorLight glossary · Streaming Multiprocessor · Tensor Core · tile · HBM · also called Tensor Cores, matrix cores, Matrix Cores