Tensor Cores
Specialised compute units in NVIDIA GPUs, designed to multiply matrices at very high speed. They do most of the work in AI.
Tensor Cores are compute units built into NVIDIA GPUs since the Volta architecture in 2017. Alongside the general-purpose CUDA cores, they can do only one thing: multiply small matrices and add up the result, in a single operation. That is precisely the calculation that comes up again and again in a neural network.
Training a model or having it generate text comes down to chaining billions of matrix multiplications. Tensor Cores process them in blocks, much faster than conventional cores. Each generation adds more compact number formats: FP16 and BF16, then FP8 with Hopper and Ada Lovelace, then FP4 with Blackwell, which is on the fifth generation. Fewer bits per number means more operations per second.
On a product page, look at the Tensor Core generation and the stated performance in TFLOPS for each precision. Always compare two cards at the same precision, and check whether the figure is given “with sparsity”, which doubles it. A consumer RTX card also has Tensor Cores: it is enough to run a model locally, whereas an H200 targets very large models and many simultaneous users.
Our AI workstations and GPU servers all rely on recent Tensor Cores, and we help you choose the right generation for your models.