Trillium
Google’s sixth-generation AI accelerator chip for training and inference.
Trillium, also known as TPU v6e, is a custom artificial intelligence chip designed by Google. It delivers high peak compute performance for transformer models and neural networks. The system connects chips in large pods to handle heavy machine learning workloads efficiently.
- Delivers 918 TFLOPs of peak BF16 compute per chip
- Contains 32 GB of high-bandwidth memory per device
- Pods scale up to 256 chips in a 2D torus topology
- Includes third-generation SparseCore for embedding tasks
- Offers 4.7x higher peak performance than TPU v5e
What it is
Trillium is the code name for the TPU v6e, a specialised processor built by Google for artificial intelligence tasks. It targets the training and serving of large language models, image generation systems, and convolutional neural networks. Each chip integrates a single TensorCore that handles matrix mathematics and vector processing. The design focuses on improving performance per cost unit compared to earlier generations. Google states that this sixth-generation device offers significantly higher peak compute power than its predecessor, the TPU v5e. It supports both integer and low-precision floating-point formats common in modern machine learning workflows.
How it came about
Google began developing its own tensor processing hardware over a decade ago to control its computing destiny. The company announced the sixth generation, codenamed Trillium, at its annual I/O event in 2024. The name references a plant species found in the Appalachian mountains, near some of Google’s data centres. This launch continues a lineage that started with the first TPU in 2015. The v6e variant appears tailored for inference workloads, following the pattern set by the v4i and v5e models. Google plans to deploy these systems in its Hypercomputer infrastructure during the fourth quarter of 2024.
Inside the chip
Each Trillium chip contains one TensorCore connected to the host server via a PCIe Gen5 interface. Inside the core, two matrix-multiply units operate on 256 by 256 arrays. A vector processing unit with 128 lanes handles other calculations, supported by cross-lane and scalar units. The chip includes 32 gigabytes of high-bandwidth memory, though only 31.25 gibibytes are available for user data due to system reserves. A feature called SparseCore accelerates operations for ranking and recommendation algorithms. Chips link together in a two-dimensional torus topology, allowing high-speed communication between neighbours in a pod.
How it was made
Specific details about the manufacturing process node and transistor count are not publicly disclosed in the provided documentation. Google has historically kept such physical specifications private for recent TPU generations. The chip connects to host machines using eight lanes of PCIe Gen5. Each pod consists of 256 chips arranged in a specific network topology. The inter-chip interconnect provides 800 gigabytes per second of bidirectional bandwidth per chip. This setup allows the entire pod to achieve massive aggregate bandwidth for reducing data across all devices during training.
Where it went
These processors reside in Google Cloud data centres and serve external customers through virtual machine instances. Configurations range from single-chip setups for testing to large multi-host slices for heavy workloads. An eight-chip slice attached to one virtual machine is optimised for inference serving. Larger topologies support distributed training of massive models. The system handles transformer architectures, text-to-image generation, and standard convolutional networks. Enterprises use them for search ranking, advertising recommendations, and general artificial intelligence research and deployment.
Versions and family
The Trillium family includes the v6e model described here. Industry observers suspect Google may release other variants, such as a v6p for heavier training tasks or a v6i for inference, similar to previous generations. The v6e shares architectural similarities with the TPU v5e but offers double the memory capacity and bandwidth. It also provides 4.7 times the peak performance per chip compared to the v5e. Previous generations like v4 and v5 supported both INT8 and BF16 formats, a capability continued in this design.
Why it matters
Google claims the Trillium chip delivers 67 percent better energy efficiency than the previous v5e generation. This improvement allows the company to offer better price-performance ratios to cloud customers despite potential price increases per device. The trend shows a significant drop in cost per unit of performance over the last decade of TPU development. By controlling its hardware stack, Google can optimise the full system from silicon to software libraries like JAX and PyTorch. This vertical integration helps maintain competitiveness against graphics processing units from other vendors.
Problems and reception
Public technical details were limited at the time of the initial announcement in mid-2024. Analysts noted missing information regarding chip size, thermal limits, and exact process nodes. Some observers questioned why Google does not publish uniform specifications for all its TPU generations. Despite the lack of granular data, the stated performance gains suggest a substantial leap in capability. The industry awaits broader availability on the cloud platform to verify real-world throughput and efficiency claims against competing accelerators.
Questions people ask about Trillium
What is the Trillium chip used for?
It is an AI accelerator designed for training and serving machine learning models. It handles transformer networks, text-to-image generation, and convolutional neural networks efficiently within Google Cloud data centres.
How much memory does each Trillium chip have?
Each chip contains 32 GB of physical high-bandwidth memory. However, users can allocate 31.25 GiB, as the system reserves a small portion for hardware reliability and runtime overhead.
How do Trillium chips connect to each other?
They use a bidirectional inter-chip interconnect with 800 GBps bandwidth per chip. In a pod, they form a two-dimensional torus topology, enabling high-speed data exchange across the system.
Is Trillium more efficient than previous Google TPUs?
Yes. Google states that Trillium offers 67 percent better energy efficiency than the TPU v5e. It also provides 4.7 times the peak performance per chip, improving overall cost-effectiveness for AI workloads.
Good reading and sources
There is no English Wikipedia article on Trillium yet, so the facts on this page come from the sources listed above: cloud.google.com/tpu/docs/v6e, nextplatform.com/ai/2024/06/10/lots-of-questions-on-googles-trillium-t. 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.
Like Trillium
All AI acceleratorsMore from Google
All of them
🇺🇸 United StatesSycamore
Google's 53-qubit superconducting quantum processorQuantum · Other · 2019
🇺🇸 United StatesTPU v2
Google AI accelerator chip for machine learning trainingAI accelerator · Other · 2019