These two chips are designed to power our custom-built supercomputers, to drive everything from cutting-edge model training and agent development, to massive inference workloads.
大多数人认为TPU主要用于加速模型训练,但作者提出TPU 8t和8i旨在支持从模型训练到推理的整个超算工作流程,挑战了TPU仅作为训练工具的传统认知。