Tensor Processing Unit
Google custom chip for neural network machine learning
The Tensor Processing Unit is a custom application-specific integrated circuit designed by Google for machine learning workloads. First used internally in 2015 and released to the cloud in 2018, it powers services like Search and Photos while offering high performance per watt for AI tasks.
- First used internally by Google in 2015
- Uses systolic array for matrix math
- Supports TensorFlow, JAX, and PyTorch
- Edge TPU consumes only 2 watts
- TPU 8t delivers 12.6 FP4 PFLOPs
What it is
The Tensor Processing Unit, or TPU, is a specialized chip built by Google to accelerate neural network machine learning. Unlike general-purpose processors, this application-specific integrated circuit focuses on high-volume, low-precision calculations. It lacks hardware for graphics tasks like texture mapping, prioritizing input and output efficiency instead. Google introduced the first version for internal use in 2015, later making it available to external developers through its cloud platform. The design supports major software frameworks including TensorFlow, JAX, and PyTorch.
How it came about
Google recruited Amir Salek in 2013 to build custom silicon capabilities for its data centers. Norman Jouppi led the architecture team, delivering the first production chip in just fifteen months. Three internal groups competed with different AI accelerator designs, but the systolic array approach won selection. The company announced the TPU in May 2016, revealing it had already operated in data centers for over a year. Broadcom co-developed the silicon, handling design translation and fabrication management with foundries like TSMC.
Inside the chip
The original TPU uses a systolic array of 8-bit multipliers driven by complex instruction set commands from a host processor. It performs matrix multiplications and convolutions without supporting rasterization. Later versions introduced floating-point support and the bfloat16 format, enabling both training and inference. The architecture scales into pods containing thousands of chips connected by high-speed networks. Recent generations separate training and inference duties, with specialized chips like the TPU 8t and 8i optimizing for distinct workload types.
How it was made
Broadcom translates Google specifications into manufacturable silicon, overseeing production at third-party foundries such as TSMC. The first generation used a 28 nanometer process with a die size under 331 square millimeters. Subsequent versions have increased memory bandwidth and core counts significantly. The eighth generation utilizes fourth-generation liquid cooling and runs on custom Arm-based Axion central processing units. This partnership covers all generations of the product line since its inception.
Where it went
Google employs these chips for RankBrain search results, Street View text processing, and photo analysis in Google Photos. They powered the AlphaGo and AlphaZero systems that mastered board games. External customers access the hardware via Google Cloud, Kaggle, and Colaboratory services. Smaller Edge TPU variants run on devices like the Coral Dev Board and Pixel smartphones for local machine learning tasks. Neocloud providers and other tech firms are also beginning to deploy them in their own data centers.
Versions and family
The product line includes eight main generations, from the initial 8-bit engine to the bifurcated TPU 8 series. The Edge TPU brings low-power acceleration to mobile and internet of things devices. Google Tensor system-on-chips integrate this technology into smartphones. Recent releases include Trillium for high performance and Ironwood for massive clustering. Each iteration improves speed, memory bandwidth, and energy efficiency compared to its predecessor.
Why it matters
The TPU established a new standard for neural network inference performance per watt in data centers. It enabled Google to process vast amounts of visual and textual data rapidly. By offering cloud access, Google allowed third parties to leverage custom AI hardware without building their own. The introduction of bfloat16 influenced numerical precision standards in machine learning. Competitors now face pressure to match the efficiency and scale of Google proprietary silicon.
Problems and reception
Singular Computing filed a patent infringement lawsuit against Google in 2019 regarding floating-point dynamic range. The case focused on claims related to bfloat16 usage exceeding standard float16 capabilities. Despite legal challenges, the technology saw widespread adoption within Google services and cloud offerings. Engineers praised the rapid development cycle and performance gains over traditional graphics processing units for specific AI workloads.

Questions people ask about Tensor Processing Unit
What is a Tensor Processing Unit?
It is a custom application-specific integrated circuit developed by Google specifically for accelerating neural network machine learning tasks. It focuses on high-volume, low-precision computations rather than general-purpose computing or graphics rendering.
Who manufactures the TPU chips?
Google designs the architecture, while Broadcom helps translate specifications into silicon. The actual manufacturing is handled by third-party foundries, primarily Taiwan Semiconductor Manufacturing Company (TSMC), covering all generations of the chip.
Can individuals buy a TPU?
Google does not sell the large data center TPUs directly to consumers. However, developers can access them through Google Cloud services. Smaller Edge TPU modules are available for purchase and integration into personal projects and devices.
What software works with TPU?
The hardware supports major machine learning frameworks including TensorFlow, JAX, and PyTorch. This allows developers to run existing models on the accelerator with minimal changes to their code, leveraging the specialized matrix multiplication engines.
Good reading and sources
- Google brings 45 teraflops tensor flow processors to its compute cloudArs Technica, Article
- Checking your browserAnandTech, Article
- Google Offers Glimpse of Third-Generation TPU ProcessorTOP500, Article
- Tearing Apart Google’s TPU 3.0 AI CoprocessorThe Next Platform, Article
- Tensor Processing Unit on WikipediaEncyclopedia
- More photos on Wikimedia CommonsPhotos
- Wikidata recordOpen data
The facts on this page come from the Wikipedia article "Tensor Processing Unit", written by its contributors and used under the CC BY-SA 4.0 licence, and from its Wikidata record (CC0). EveryChip rewrote the text in its own words, with AI assistance, and compares every page with the article to make sure no passage is copied. It may be reused under the same CC BY-SA 4.0 licence with a link back.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.
Like Tensor Processing Unit
All AI acceleratorsMore from Google
All of them
🇺🇸 United StatesSycamore
Google's 53-qubit superconducting quantum processorQuantum · Other · 2019
🇺🇸 United StatesTPU v2
Google AI accelerator chip for machine learning trainingAI accelerator · Other · 2019
🇺🇸 United StatesTrillium
Google’s sixth-generation AI accelerator chip for training and inference.AI accelerator · Tensor design · 2024