Paper reference: In-Datacenter Performance Analysis of a Tensor Processing Unit”
Application
Using floating point (16bit or 32bit) for NN (Neural Network) training, then a step called quantization transforms floating-point numbers into narrow integers–often just 8 bits–which are usually good enough for inference.
MLP(Multi-layer Perceptions), CNN(Convolutional Neural Netowrks), and RNN(Recurrent Neural Networks), these three types of NN represent 95% of NN inference workload in Google datacenter. Therefore, the TPU mainly focus on them.



As we can see, CNNs are usually dense-computing NN, which are better for TPU.

TPU has 25 times as many MACs (Multiply and Accumulate) and 3.5 times as much on-chip memory as the K80 GPU.
Architecture
The TPU was designed to be a coprocessor on the PCIe I/O bus, more like FPU(floating-poin unit) than it is to a GPU.


The parameters of NN model (weights) comes from off-chip memory (8G DDR3 DRAM) to Weight FIFO, and then flow into MMU(Matrix Multiply Unit). The request (sample need to be inference) comes from PCIe to Unified Buffer, and also flow into MMU finally.
Even the “Activation” and “Pooling” algorithm in CNN have been fixed into hardware.

The MMU contains 256×256 MACs that can perform 8-bit multiply-and-adds on signed or unsigned integers.


According to this Floor Plan, we can imaging that UB and MMU might cost most energy of TPU.

TPU instructions follow the CISC tradition and only has about a dozen instructions, include “Read_Host_Memory”, “Read_Weights”, “MatrixMultiply”, “Activate” etc. Recalling how many codes we need to write to implement a effective Activation function, then we could conceive the speed of using only one “Activate” instruction in TPU.
This paper said TPU is a type of Systolic Array. But what is Systolic Array? Here is the explain: A systolic array is a network of processors that rhythmically compute and pass data through the system.
Performance
There are lot of tables and diagrams which show the top-rate performance of TPU. Although the TPU is fast, it also depend on the computing-density of applications. The CNNs are most computing-dense NN, so it gains most speed(or TeraOps per second) from TPU:



In this paper, it didn’t explain why the GPU is slower than TPU in inference operation. The only sentence about this topic is in “8 Discussion”: “GPUs have traditionally been seen as high-throughput architectures that reply on high-bandwidth DRAM and thousands of threads to achieve their goals”. Actually, I think this is not a serious explain.
The interesting thing is, after Google publish this paper, the CEO of Nvidia – Jensen Huang, wrote a blog to gently appeal a fact: the state-of-the-art GPU (Tesla P40) can inference faster than TPU. The war between different giants of Deep learning is just beginning.