News
IndustryTSMC’s grip on advanced chipmaking will face competitionIndustryCadence joins European Chip Design Platform to support semiconductor startups and SMEsMaking chipsMarc Hijink on ASML’s Culture, Supply Chain, and GeopoliticsProcessorsIntel's post-Nova Lake CPUs take shape as Razor Lake desktop and laptop variants surfaceAI chipsHow Google’s New AI Chip Software Helps It Battle NvidiaIndiaMore IC Packaging Plants Emerge In IndiaIndustryUS Tightens China AI Chip Access as Nvidia, TSMC and Semiconductor Stocks SlideDesignAxelera AI moves into HPC with TitaniaMaking chipsChinese envoy says chip equipment export curbs threaten EU-China trade thawIndiaSEMICON India 2026 Day 2 Sees 25 MoUs, Strategic Collaborations Across Semiconductor Value ChainIndustryTSMC’s grip on advanced chipmaking will face competitionIndustryCadence joins European Chip Design Platform to support semiconductor startups and SMEsMaking chipsMarc Hijink on ASML’s Culture, Supply Chain, and GeopoliticsProcessorsIntel's post-Nova Lake CPUs take shape as Razor Lake desktop and laptop variants surfaceAI chipsHow Google’s New AI Chip Software Helps It Battle NvidiaIndiaMore IC Packaging Plants Emerge In IndiaIndustryUS Tightens China AI Chip Access as Nvidia, TSMC and Semiconductor Stocks SlideDesignAxelera AI moves into HPC with TitaniaMaking chipsChinese envoy says chip equipment export curbs threaten EU-China trade thawIndiaSEMICON India 2026 Day 2 Sees 25 MoUs, Strategic Collaborations Across Semiconductor Value Chain
EveryChip
TPU v2

TPU v2

Google AI accelerator chip for machine learning training

The TPU v2 is an application-specific integrated circuit designed by Google to accelerate machine learning model training. Released in public beta in May 2019, it connects multiple chips into large pods for high-performance cloud computing tasks.

Share on WhatsApp

  • Each device has 4 chips and 8 cores
  • Full Pod contains 256 chips and 512 cores
  • Released in public beta in May 2019
  • Used by eBay, Lyft, and Recursion Pharmaceuticals
  • Reduces training time from days to hours

What it is

The TPU v2 is a specialised hardware accelerator built by Google for artificial intelligence workloads. It functions as an application-specific integrated circuit, meaning its design is fixed to perform specific mathematical operations required for deep neural networks efficiently. Unlike general-purpose processors, this chip prioritises speed and energy efficiency for training machine learning models. Google offers these units through its cloud platform, allowing developers to access significant computational power without managing physical hardware. The system supports various configurations, from single devices to large interconnected pods.

How it came about

Google introduced the Cloud TPU v2 Pods during the Google I/O 2019 conference in Mountain View, California. This launch marked the first time these powerful clusters were available in public beta, having previously been restricted to a private alpha testing phase. The release aimed to provide machine learning engineers with faster deployment options for complex models. By making the technology publicly accessible, Google sought to compete with other tech giants developing their own AI silicon. The announcement coincided with growing industry interest in dedicated AI hardware beyond traditional graphics processors.

Inside the chip

Each TPU v2 device contains four individual chips and eight processing cores. These units can be linked together to form a Pod, which acts as a supercomputer for cloud-based tasks. A full Pod consists of 64 devices, totaling 256 chips and 512 TensorCores. High-speed reconfigurable links connect these components to facilitate rapid data exchange during model training. Users can create smaller slices of this infrastructure, such as a v2-8 or v2-128 configuration, depending on their computational needs. This modular design allows for scalable performance tailored to specific workload sizes.

How it was made

Specific details regarding the manufacturing process node, transistor count, or physical packaging are not provided in the available documentation. The chips are designed by Google as custom application-specific integrated circuits. The focus remains on the logical architecture and cloud integration rather than the physical fabrication metrics.

Where it went

Organisations use TPU v2 Pods to train large production models significantly faster than on-premises clusters. Companies such as eBay, Lyft, Two Sigma, and Recursion Pharmaceuticals have adopted the technology. Recursion Pharmaceuticals reported reducing model training time from over 24 hours to just 15 minutes for testing synthesized molecules. The hardware supports frequent model retraining and improves accuracy for deep learning tasks. Developers can iterate on their algorithms in minutes, accelerating the path from experimental code to deployed applications in fields like healthcare and finance.

Versions and family

The TPU v2 family includes several supported types, ranging from v2-8 to v2-512. These variants differ by the number of TensorCores and chips allocated to the user. The v2-8 represents a single device, while larger slices like v2-128 offer 128 cores. The entire Pod configuration provides the maximum capacity with 512 cores. Google also released the TPU v3, which features upgraded hardware and higher pricing but serves similar functions. The v2 API is no longer under active development, receiving only security updates and bug fixes.

Why it matters

The public availability of TPU Pods represented a shift in the accelerated server market, which had been dominated by graphics processing units. Analysts noted that while GPUs offer flexibility and a mature software ecosystem, dedicated AI ASICs provide superior performance for specific training tasks. Google’s move encouraged other vendors to develop their own AI-centric chips. The technology demonstrated that cloud-based supercomputers could dramatically reduce the time and cost associated with training complex neural networks, influencing how enterprises approach large-scale machine learning projects.

Problems and reception

Industry observers described the TPU Pods as extremely powerful compared to other available options at the time. Experts highlighted the benefits of application-specific circuits, including speed and low energy consumption, while noting their lack of flexibility compared to programmable GPUs. Users reported substantial improvements in training efficiency, with some achieving order-of-magnitude reductions in processing time. The reception focused on the practical advantages for data scientists who needed to handle large workloads without the overhead of managing physical infrastructure.

Questions people ask about TPU v2

What is the core count of a single TPU v2 device?

A single TPU v2 device contains four individual chips and a total of eight processing cores. These cores are designed specifically for tensor operations used in machine learning.

How many chips are in a full TPU v2 Pod?

A complete Cloud TPU v2 Pod consists of 64 devices, which amounts to a total of 256 TPU chips. These chips provide 512 TensorCores for parallel processing.

When did Google make TPU v2 Pods publicly available?

Google announced the public beta availability of Cloud TPU v2 Pods on May 7, 2019. This launch occurred during the Google I/O conference in Mountain View, California.

Can users rent partial TPU v2 Pods?

Yes, users can create slices of a Pod using specific accelerator types. Options include v2-8, v2-32, v2-128, and v2-256, allowing customers to pay for only the resources they need.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on TPU v2 yet, so the facts on this page come from the sources listed above: cloud.google.com/tpu/docs/v2, techtarget.com/ai/news/252462950/Googles-Cloud-TPU-v2-v3-Pods-accelera. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryChip wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    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.

    Photo credits on this page

    The small pictures above come from Wikimedia Commons under free licences. Some are cropped to fit.

    • Sycamore: Google ( https://www.youtube.com/channel/UCK8sQmJBp8GCxrOtXWBpyEA ), CC BY 3.0, source
    • Tensor Processing Unit: Norman P. Jouppi, George Kurian, Sheng Li, Peter Ma, Rahul Nagarajan,, CC BY 4.0, source

    The weekly digest

    Every Monday: the chip of the day, new pages on EveryChip, and the conferences and anniversaries of the week ahead.