What is a TPU?

A tensor processing unit (TPU) is a specialized integrated circuit developed by Google for neural network machine learning. The TPU is a critical component of Google’s AI Supercomputer which enables the company to develop large language models, that are spurring up the current AI revolution.

Understanding TPUs

Tensor processing units – also known as TensorFlow processing units – are machine learning (ML) accelerators in the form of specialized integrated circuits. TPUs were created by Google to handle the neural network processing of TensorFlow – the company’s free and open-source software library for ML and AI purposes.

More specifically, Google introduced TPUs in 2016 to deal with matrix multiplication operations in neural network training. They can be accessed in two different forms. The first, Cloud TPU, is offered as part of the Google Cloud Platform and hosted in the company’s data centers.

The second, Edge TPU, followed in 2018 and forms part of a custom development kit that is used to build specific applications. Edge TPU is an application-specific integrated chip (ASIC) created to run ML models in edge computing contexts, and it was made available to developers with products within the Coral brand in January 2019.

The key components of a TPU

Some of the components (and indeed vocabulary) associated with a TPU include:

Tensors – fundamental units of a neural network that store data (such as the weights of a node) in a row and column format. Tensors are used to perform basic mathematical calculations like addition and matrix multiplication.
FLOPs – floating-point operations per second (FLOPs) are used to measure the speed of the computer operation. In Google TPUs, this unit of measurement is called brain floating-point format – or bfloat16 for short – which is placed within systolic arrays to accelerate training. A higher FLOPs range is associated with higher processing power.
Systolic array – the collection of sensors responsible for executing computations and distributing the results across the system. These processing elements are grouped such that they are ideal for parallel processing.

Where are TPUs used?

Tensor processing units are used in DeepMInd’s AlphaGo – a computer program that beat the world’s best human player at the Chinese board game Go. TPUs are also present in the AlphaZero program which in addition to Go has mastered the games of chess and shogi.

TPUs have also been used to power many applications at Google itself. One example is RankBrain – the core component of Google’s algorithm which employs machine learning to determine the most relevant search results. Another is in Google Photos, where a TPU can process over 100 million photos per day. TPUs were also utilized to improve the quality and accuracy of maps of navigation within Street View.

Otherwise, tensor processing units are the most effective when models need to rely on matrix computations (such as the recommendation systems for search engines we mentioned above). They are also useful when AI needs to train new models from scratch with vast amounts of data.

Key takeaways:

A Tensor Processing Unit (TPU) is a specialized integrated circuit developed by Google for neural network machine learning.
Google introduced TPUs in 2016 specifically to deal with matrix multiplication operations in neural network training. They can be accessed in two different forms: Cloud TPU and Edge TPU.
Tensor processing units are used in DeepMInd’s AlphaGo and AlphaZero game-based programs. They have also been used to power many applications at Google itself such as Google Photos, RankBrain, and Street View.