Loading…
Glossary term
A specialized unit in modern GPUs that accelerates the matrix math at the heart of neural networks.
Tensor cores are units inside modern GPUs built to multiply small blocks of numbers very quickly, often in lower precision. Because neural networks are mostly matrix multiplication, they deliver a large speed-up.
They are what make formats such as bfloat16 and 8-bit so effective.
They are a main reason recent GPUs are far faster at AI work than at general computing.