(TPU)In-datacenter performance analysis of a tensor processing unit

Rather than be tightly integrated with a CPU, to reduce the chances of delaying deployment, the TPU was designed to be a coprocessor on the PCIe I/O bus, allowing it to plug into existing servers just as a GPU does.

为了减少延迟部署的可能性，TPU 不是与 CPU 紧密集成，而是被设计为 PCIe I/O 总线上的协处理器，允许它像 GPU 一样插入现有服务器。

TPU是挂在PCLe总线上，与CPU是并行关系。

Moreover, to simplify hardware design and debugging, the host server sends TPU instructions for it to execute rather than the TPU fetching them itself. Hence, the TPU is closer in spirit to an FPU (floating-point unit) coprocessor than it is to a GPU.

此外，为了简化硬件设计和调试，主机服务器发送 TPU 指令供其执行，而不是 TPU 自己获取它们。因此，与 GPU 相比，TPU 在精神上更接近 FPU（浮点单元）协处理器。

 CPU向TPU发送控制指令，TPU执行。