TPU v3
Google AI accelerator chip with water cooling
The TPU v3 is a third-generation artificial intelligence accelerator designed by Google. It features two TensorCores per chip and requires water cooling to manage heat. The system supports large-scale machine learning training through high-bandwidth interconnects.
- Each chip has two TensorCores with matrix-multiply units
- Requires water cooling due to high power density
- Full pod contains 1024 chips in a 2D torus layout
- Provides 32 GiB of HBM2 memory per chip
- Peak compute is 123 teraflops per chip in bf16
What it is
The TPU v3 is a specialised integrated circuit built for artificial intelligence workloads. Google designed this third-generation hardware to accelerate deep neural network training. Each physical chip holds two processing units called TensorCores. These cores handle matrix mathematics, vector operations, and scalar tasks. The system delivers high computational throughput measured in teraflops. It uses high-bandwidth memory to feed data quickly to the compute units. This design targets large-scale models that require significant parallel processing power.
How it came about
Google introduced the first generation of its Tensor Processing Units in 2016. The company had already used them internally for over a year before that announcement. The second generation added support for single-precision floating-point numbers and more memory. The third generation continued this trend of increasing performance density. Google CEO Sundar Pichai stated that these new pods are eight times more powerful than the previous version. This jump in capability required a change in thermal management strategies.
Inside the chip
Each v3 chip contains two TensorCores. Every TensorCore includes two matrix-multiply units, a vector unit, and a scalar unit. The chip connects to 32 gigabytes of HBM2 memory with a bandwidth of 900 GBps. Multiple chips link together in a two-dimensional torus topology. This structure allows efficient data movement across the entire pod. A full pod consists of 1024 chips. The interconnect provides substantial all-reduce bandwidth for synchronising training steps. This architecture helps reduce the need to recompute intermediate values during model training.
How it was made
Specific details about the manufacturing process node are not provided in the available documentation. Previous generations used TSMC facilities, but the foundry for this version is not explicitly named here. The physical design incorporates a copper-plated cooler system. Water runs through pipes connected to four cooling plates on each unit. This liquid cooling solution addresses the high heat output generated by the dense compute clusters. Power consumption ranges from a minimum of 123 watts to a maximum of 262 watts per chip.
Where it went
These accelerators serve data centres running large machine learning models. They support configurations ranging from single chips to pods with thousands of units. Users can deploy them via Google Kubernetes Engine or the Cloud TPU API. The hardware handles compute-bound models effectively due to increased flops per core. It also assists with memory-bound tasks by providing larger capacity than earlier versions. Researchers use them to train deeper networks like ResNet with larger batch sizes that did not fit on prior hardware.
Versions and family
Google offers several slice sizes for this accelerator type. Options include v3-8, v3-32, v3-128, v3-256, v3-512, v3-1024, and v3-2048. The number indicates the count of TensorCores in the configuration. A single v3-8 device contains one chip with two TensorCores. Larger slices combine multiple chips into a unified resource. The v3-1024 configuration represents a full pod with 1024 chips. All variants share the same underlying architectural principles and memory specifications.
Why it matters
The introduction of water cooling marked a shift in how cloud providers manage high-density AI hardware. Previous TPUs relied on air cooling, but the thermal density of the v3 required liquid solutions. This move enabled higher performance per rack unit. The system allows for training models that were previously constrained by memory limits on v2 hardware. By reducing recomputation needs, it improves efficiency for certain types of neural network training workflows.
Questions people ask about TPU v3
Does the TPU v3 use air or water cooling?
The TPU v3 requires water cooling. Google implemented a copper-plated cooler system with water pipes running through four cooling plates. This change was necessary because the chip generates too much heat for traditional air cooling methods.
How much memory does each TPU v3 chip have?
Each individual chip contains 32 gigabytes of HBM2 memory. The memory bandwidth is rated at 900 GBps. This capacity is double that of the previous generation, allowing larger models to fit in memory.
What is the largest configuration available for TPU v3?
The largest standard configuration is the v3-2048 slice. A full TPU v3 Pod consists of 1024 chips interconnected together. These pods deliver peak compute performance of 126 petaflops in bf16 precision.
Good reading and sources
There is no English Wikipedia article on TPU v3 yet, so the facts on this page come from the sources listed above: cloud.google.com/tpu/docs/v3, extremetech.com/extreme/269008-google-announces-8x-faster-tpu-3-0-for-. 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.