A processing unit in an NVIDIA GPU that accelerates AI neural network processing and high-performance computing (HPC). There are typically from 300 to 600 Tensor cores in a GPU, and they compute values on Tensor arrays in parallel. The multidimensional Tensor array is turned into many two-dimensional matrices for the actual computations.
Not Just NVIDIA
Tensor Core is NVIDIA's product, but other companies make math acceleration hardware. Google's TPU Matrix Units (MXUs), Apple's Matrix Coprocessor (AMX) and AMD's Matrix Cores are examples.
Multiply and Add
A new array is created from two arrays by quickly multiplying and adding 16- or 32-bit floating point values in parallel. Optimized for mixed precision, Tensor cores work with both floating point and integer math. See
TF32.
NVIDIA Tensor vs. CUDA
NVIDIA's CUDA cores were designed for general-purpose parallel processing, and they are used in AI applications. However, Tensor cores were designed for AI deep learning, and both CUDA and Tensor cores are found in NVIDIA GPUs. See
CUDA core,
matrix multiplication,
GPU,
TPU,
neural network,
Tensor Processing Unit and
Tensor chip.