tensor core
also: tensor unit, matrix engine
A specialized hardware processing unit in NVIDIA GPUs designed to accelerate matrix multiplication and deep learning operations. Tensor cores perform low-precision mathematical operations (like matrix-matrix multiplication) much faster than general-purpose GPU cores.
Tensor cores are dedicated compute elements found in NVIDIA's modern GPUs (starting with the Volta architecture). They perform mixed-precision calculations optimized for machine learning workloads, particularly matrix operations that are fundamental to neural network training and inference.
Unlike traditional GPU cores that process scalar or vector operations sequentially, tensor cores operate on entire small matrices in a single clock cycle. For example, a tensor core can multiply two 4×4 matrices and accumulate the result into another 4×4 matrix in one operation, dramatically accelerating deep learning frameworks like TensorFlow and PyTorch.
Tensor cores support various precision levels: FP32 (full precision), TF32, FP16 (half precision), and INT8 (integer). This flexibility allows developers to trade numerical precision for speed and memory efficiency, which is crucial for training large models. Modern GPU models like the H100 contain thousands of tensor cores, making them essential accelerators for AI and scientific computing workloads on Linux systems.