TPU v5p
Google AI accelerator chip for large-scale model training
The TPU v5p is an artificial intelligence accelerator designed by Google for its Cloud platform. It features high-bandwidth memory and a 3D torus interconnect to support massive language models like Gemini.
- Each chip has 96 GiB of high-bandwidth memory
- Pods contain 8960 interconnected chips
- Supports 3D torus network topology
- Trains LLMs 2.8x faster than TPU v4
- Used to train the Gemini AI model
What it is
The TPU v5p is a specialised integrated circuit built for machine learning workloads. Google designed this chip to handle the intense computational demands of training large language models. It serves as the hardware backbone for the company’s generative AI services, including the Gemini model family. Each unit contains dedicated matrix multiplication engines and vector processing units. The system scales into massive clusters called Pods, allowing thousands of chips to work in unison on single tasks.
How it came about
Google introduced the v5p alongside its Gemini AI models. This launch marked a shift toward more frequent hardware disclosures by the company. The design aims to keep Google competitive in the generative AI sector against rivals like NVIDIA. It follows the earlier TPU v4 and the efficiency-focused v5e. The company stated that this generation offers significant speed improvements for training large neural networks compared to previous iterations.
Inside the chip
Each chip houses two TensorCores and four SparseCores. A TensorCore includes four Matrix Multiply Units and a vector unit. The chip connects to 96 gibibytes of high-bandwidth memory, though software reserves some for reliability. Inter-chip communication occurs via a bidirectional link with 1200 gigabytes per second of bandwidth. Chips form a three-dimensional torus network topology. This structure allows data to wrap around the edges of the grid, improving communication symmetry for certain model shapes.
How it was made
Google groups 8960 chips into a single Pod. These Pods use reconfigurable high-speed links to connect the silicon. Optical circuit switches manage connections between cubes of chips, while copper links handle internal cube traffic. The system supports twisted torus topologies for specific slice dimensions. This configuration boosts bisection bandwidth by up to 70 percent for asymmetric shapes. Fault tolerance features allow the network to route around optical link failures automatically.
Where it went
Data centres use these accelerators for training and running large language models. Google Cloud customers access the hardware through virtual machines with 208 virtual CPUs. The system supports single-slice training on up to 6144 chips. Larger jobs can scale to over 18000 chips using Multislice technology. Workloads benefit from the high memory bandwidth and fast interconnects. Embedding dense models train nearly twice as fast as on prior generations.
Versions and family
The v5p sits above the v5e in performance and capability. It replaces the TPU v4 as the flagship training accelerator. Supported topologies range from small 2x2x1 slices to large 16x16x24 configurations. Twisted topologies are available for slices where dimensions meet specific size ratios. The full Pod contains 140 cubes of chips. Machine types like ct5p-hightpu-4t provide access to four-chip VMs for developers.
Why it matters
Google claims the v5p delivers 2.3 times better performance per dollar than the v4. Training speeds for large language models reportedly increased by a factor of 2.8. This hardware enables the creation of larger and more complex AI models. It helps Google retain engineering talent in a competitive market. The exclusive availability on Google Cloud reinforces the company’s ecosystem lock-in strategy for enterprise AI users.
Problems and reception
Industry observers note the chip is only available via Google Cloud. This restriction limits its appeal for on-premise deployments. Some critics argue that cloud lock-in remains a serious concern for businesses. However, the frequent disclosure of hardware details was welcomed by technical communities. The performance gains for specific AI workloads are seen as a strong counter to competing GPU offerings.
Questions people ask about TPU v5p
How much memory does each TPU v5p chip have?
Each chip contains 96 gibibytes of physical high-bandwidth memory. The system reserves a small portion for reliability, leaving 95 gibibytes available for user applications and machine learning frameworks.
What is the maximum size of a TPU v5p Pod?
A full TPU v5p Pod consists of 8960 chips. These are organised into 140 cubes. The largest single slice supported for training uses 6144 chips, but multislice configurations can scale further.
How does the twisted torus topology help?
Twisted topologies increase bisection bandwidth by up to 70 percent for certain slice shapes. They convert asymmetric grids into symmetric ones, which improves performance for workloads using global communication patterns.
Can I buy a TPU v5p for my own server?
No. The TPU v5p is exclusively available through Google Cloud. It is not sold as a standalone component for on-premise servers or other cloud platforms, requiring users to adopt Google’s infrastructure.
Good reading and sources
There is no English Wikipedia article on TPU v5p yet, so the facts on this page come from the sources listed above: cloud.google.com/tpu/docs/v5p, 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.