Question

a)

l) Three architectural reasons why the TPU succeeded are?

ll) Describe briefly the hardware design of the TPU, clearly identifying the physical connector used in a server, the bus over which TPU instructions are sent, and the TPU’s relationship to the CPU.

Answer 1

a.

I.

Tensor Processing Units:

It’s used in TensorFlow framework.

The three advantages why TPU is a success:

1. It’s architecture reduces the time taken to bring the accuracy while training models which use neural networks.

2. It performs faster than the GPU and CPU developed before this. While GPU wasn’t totally based on machine learning, TPU is specifically designed for that purpose.

3. There’s no need to write a lot of code, a few lines would do the job.

Hence these are a few of the reasons why TPU succeeded as they were able to improve.

II.

Hardware Design:

The DDR3 chips are used with three interfaces. Unified Buffers are used as registers, Matrix Multiplier Unit is used to perform matrix operations. It’s domain specific architecture used the tensorflow server which connected all the queries required by the user such as speech recognition, searching as TPU is developed by Google. It reduced the precision of the chip from floating units to int units by the process of quantization. The usage of memory to perform larger operations was costly hence this chip uses Matrix Multiplier Unit in which symbolic array is used, where it helps to perform all the calculations without accessing the memory. As much a 250,000 operations are performed per clock cycle. Whenever google translate or such operations are used, the TPU supports this till some extent. There are many TPU instructions such as Activate, Convolve, Read_Host_Memory, Write_Host_Memory, etc all these are sent by the Peripheral component interconnect express(PCIe) Gen3 x16 bus. The relationship of TPU with CPU is that CPU performs sequential execution while TPU performs as a whole. Simple models use CPU, while the models which take a longer time to get trained use TPU as it reduces this time. TPU is a co-processor. Both CPU, TPU perform operations by CPU takes in tens while TPU takes in 250,000.

comment for any doubts!