Tensor Processing Unit / TPU
Developed by Google, specially designed accelerator chip for artificial intelligence workloads.
Tensör Trading Unit (TPU) is a type of integrated circuit (ASIC) which is specifically designed from scratch for deep learning workloads by Google. In contrast to general purpose GPUs, TPU is optimized to make the base of nerve networks with the highest efficiency of matrix palpitation and collection; it uses a architecture called "stolic array" in the core and thus reduces unnecessary memory access by constantly assigning between data and processors. This specialization allows TPUs to deliver higher energy efficiency and speed compared to GPUs in certain deep learning tasks.
TPUs are offered only on Google Cloud and are used intensively in the training of Google’s own major models; frames such as TensorFlow and JAX have been developed directly compatible with TPU. The emergence of TPUs has shown that the artificial intelligence hardware market is not in the monopoly of NVIDIA GPUs, companies can achieve cost and performance advantages by designing special chip on their own business loads, which then inspired similar initiatives such as Amazon’s Trainium. TPU generations continue to develop to offer more calculation strength and memory bandwidth with each new version.
TPUs are offered only on Google Cloud and are used intensively in the training of Google’s own major models; frames such as TensorFlow and JAX have been developed directly compatible with TPU. The emergence of TPUs has shown that the artificial intelligence hardware market is not in the monopoly of NVIDIA GPUs, companies can achieve cost and performance advantages by designing special chip on their own business loads, which then inspired similar initiatives such as Amazon’s Trainium. TPU generations continue to develop to offer more calculation strength and memory bandwidth with each new version.