# TPU v1

TPU v1 is Google's first-generation Tensor Processing Unit, a custom ASIC for neural network inference. Announced in 2016, it delivers 15-30x higher performance and 30-80x better performance-per-watt than contemporary CPUs and GPUs for machine learning workloads.

The TPU v1 is the first-generation Tensor Processing Unit, a custom application-specific integrated circuit (ASIC) developed by Google for accelerating neural network inference. Introduced internally in 2015 and publicly announced in May 2016 at Google I/O, it was designed specifically for the TensorFlow framework. The chip is a systolic array matrix multiplication engine optimized for high-volume, low-precision computation, primarily using 8-bit integers, and was deployed across Google's data centers to power services such as search, Street View, and Photos.

Unlike general-purpose processors, the TPU v1 was built to handle the specific computational patterns of [neural networks](https://www.wikiprompt.org/wiki/neural-network) at scale. Its architecture, which shares a genealogy with earlier systolic array systems like the WARP, was selected from among three competing AI accelerator designs at Google. The development was led by Norman P. Jouppi as tech lead and principal architect, with Amir Salek founding Google's custom silicon group in 2013. The chip was brought from design to production in just 15 months, a rapid timeline for a custom ASIC.

## Architecture and Specifications

The TPU v1 is an 8-bit matrix multiplication engine driven by complex instruction set computer (CISC) instructions from a host processor over a PCIe 3.0 bus. It is manufactured on a 28 nm process with a die size of no more than 331 mm². The chip operates at a clock speed of 700 MHz and has a thermal design power of 28-40 W, making it highly energy-efficient for its intended workloads.

Key components include 28 MiB of on-chip memory and 4 MiB of 32-bit accumulators that collect results from a 256×256 systolic array of 8-bit multipliers. The package contains 8 GiB of dual-channel 2133 MHz DDR3 SDRAM, providing 34 GB/s of memory bandwidth. Instructions handle data transfer to and from the host, perform matrix multiplications or convolutions, and apply activation functions, covering the core operations needed for inference.

## Performance and Deployment

In a seminal 2017 paper titled "In-Datacenter Performance Analysis of a Tensor Processing Unit," presented at the 44th International Symposium on Computer Architecture (ISCA 2017), Jouppi and colleagues demonstrated that the TPU v1 achieved 15-30 times higher performance and 30-80 times higher performance-per-watt than contemporary CPUs and GPUs. This established the chip as a foundational platform for neural network inference at scale.

Google deployed TPUs across its production services. For [Google](https://www.wikiprompt.org/wiki/google-cloud) Street View text processing, the chip found all text in the database in less than five days. In Google Photos, a single TPU could process over 100 million photos per day. The chip also powered RankBrain, which Google uses to improve search results, and was used in the AlphaGo versus Lee Sedol Go matches and the AlphaZero system.

## Comparison with GPUs and CPUs

TPUs are designed for high-volume, low-precision computation with more input/output operations per joule, without hardware for rasterization or texture mapping. This makes them well suited for convolutional neural networks (CNNs), which dominate many inference tasks. In contrast, GPUs have benefits for some fully connected neural networks, and CPUs can have advantages for recurrent neural networks (RNNs).

Different processor types serve different machine learning models. For transformer-based neural networks used in large language models, TPUs are often used for inference while GPUs are used for training. The TPU ASICs are mounted in heatsink assemblies that can fit in hard drive slots within data center racks, facilitating dense deployment.

## Legacy and Impact

The TPU v1's success paved the way for subsequent generations, including the TPU v2, which introduced bfloat16 floating-point support for training, and later versions. Broadcom has served as a co-developer, translating Google's architecture into manufacturable silicon and managing fabrication through foundries like [TSMC](https://www.wikiprompt.org/wiki/tsmc). The chip's design influenced the broader AI accelerator landscape, with other vendors developing similar products for embedded and robotics markets.

Google made TPUs available to third parties through its Cloud TPU service on [Google Cloud Platform](https://www.wikiprompt.org/wiki/google-cloud) in 2018, and through notebook-based services like Kaggle and Colaboratory. The TPU v1 remains a landmark in [artificial intelligence](https://www.wikiprompt.org/wiki/artificial-intelligence) hardware, demonstrating the value of domain-specific processors for [machine learning](https://www.wikiprompt.org/wiki/machine-learning) workloads.

---
Source: https://www.wikiprompt.org/wiki/tpu-v1
License: CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/)
Last updated: 2026-09-13T03:59:57.770504+00:00