TPU v5e
Google AI accelerator for training and inference workloads
The TPU v5e is an artificial intelligence chip designed by Google for machine learning tasks. It supports both model training and inference, offering specific performance metrics for bf16 and Int8 operations within cloud environments.
- Each chip delivers 197 TFLOPs of bf16 peak compute
- Contains 16 GiB of high-bandwidth memory per chip
- Supports training scales up to 256 chips per job
- Uses a 2D torus interconnect topology for chips
What it is
The TPU v5e is a specialised integrated circuit developed by Google for accelerating artificial intelligence workloads. It functions as a combined product for both training machine learning models and running inference tasks. The chip is designed to operate within Google Cloud infrastructure, supporting single and multi-host configurations. Each unit contains a single TensorCore, which handles the primary computational duties. This architecture allows the device to manage large-scale data processing required for modern generative AI and other deep learning applications.
Inside the chip
Each TPU v5e chip houses one TensorCore. This core includes four matrix-multiply units, known as MXUs, along with a vector unit and a scalar unit. The system uses high-bandwidth memory, providing 16 GiB of capacity per chip with a bandwidth of 819 GBps. Chips connect via a two-dimensional torus topology, allowing 400 GBps of bidirectional inter-chip communication. For virtual machine setups, the chips are arranged in slices of one, four, or eight. In larger eight-chip configurations, the system uses non-uniform memory access nodes to optimise communication between the central processing units and the accelerator chips.
Where it went
This hardware is deployed in data centres for cloud-based computing. It supports training jobs that prioritise throughput and availability, as well as serving jobs focused on low latency. Users can provision these resources through the Cloud TPU API or command line interfaces. The system supports single-host inference for up to eight chips and multi-host inference using specific software tools. Training configurations can scale up to 256 chips in various topological shapes, such as 4x4 or 16x16 grids, depending on the computational requirements of the machine learning model.
Questions people ask about TPU v5e
What is the primary function of the TPU v5e?
The TPU v5e is an AI accelerator designed for both training machine learning models and performing inference. It is built to handle large-scale computational tasks within Google Cloud, optimising for either throughput during training or latency during serving.
How much memory does each TPU v5e chip have?
Each individual chip contains 16 GiB of high-bandwidth memory (HBM). The bandwidth for this memory is rated at 819 GBps per chip, allowing for rapid data movement during complex mathematical operations.
What is the internal structure of the TPU v5e chip?
Each chip contains a single TensorCore. This core is composed of four matrix-multiply units, one vector unit, and one scalar unit. These components work together to process the tensor operations required for deep learning algorithms.
Good reading and sources
There is no English Wikipedia article on TPU v5e yet, so the facts on this page come from the sources listed above: cloud.google.com/tpu/docs/v5e, servethehome.com/google-tpu-v5p-ai-chip-launches-alongside-gemini/. 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.