News
EveryChip
Ironwood

Ironwood

Google's seventh-generation AI accelerator for inference workloads

Ironwood is Google's seventh-generation Tensor Processing Unit, designed specifically for AI inference and training. Announced in 2025, it offers high compute power and energy efficiency for large language models and scientific workloads on Google Cloud.

Share on WhatsApp

  • Scales up to 9,216 chips in a single pod configuration
  • Delivers 192 GB of HBM capacity per individual chip
  • First TPU to natively support FP8 calculations in tensor cores
  • Offers 42.5 exaflops of total compute power at full scale
  • Twice the performance per watt compared to previous Trillium TPU

What it is

Ironwood is the seventh generation of Google’s custom Tensor Processing Units. It serves as an artificial intelligence accelerator built for data centres. The company states it is their most powerful and energy-efficient TPU to date. Unlike previous generations focused primarily on training, this chip targets the age of inference. It handles complex tasks like large language model reasoning and mixture of experts architectures. The design supports both training and serving workloads at massive scale.

How it came about

Google introduced Ironwood at the Google Cloud Next 25 event in Las Vegas. This launch marks a shift from responsive AI models to proactive insight generation. The company has developed TPUs for over a decade to support its own services. They began offering these chips to external customers via Google Cloud seven years ago. Ironwood represents the latest step in this internal development programme. It aims to meet the rising computational demands of modern generative AI applications.

Inside the chip

The architecture features two compute chiplets within each package, suggesting a non-monolithic design. It includes eight banks of high bandwidth memory, totaling 192 gigabytes per chip. This capacity is six times that of the previous Trillium generation. The chip supports FP8 calculations natively in its tensor cores, a first for the series. It also contains third-generation SparseCore accelerators for processing large embeddings. These units help with financial and scientific calculations beyond standard AI tasks. Inter-chip interconnect bandwidth reaches 1.2 terabytes per second bidirectionally.

How it was made

Google designs these accelerators in-house for use in its cloud infrastructure. The packaging includes central chiplets above compute dies and additional I/O dies. Eight stacks of memory surround the compute elements. The system relies on advanced liquid cooling solutions to manage heat. This approach allows the hardware to sustain high performance under heavy loads. The company claims this cooling method is twice as effective as standard air cooling for these workloads.

Where it went

Ironwood powers Google Cloud AI Hypercomputer configurations. Customers can access clusters of 256 or 9,216 chips. These pods deliver up to 42.5 exaflops of compute power. The hardware supports demanding models like Gemini 2.5 and AlphaFold. Developers use the Pathways software stack to distribute tasks across thousands of chips. This setup enables efficient handling of super large dense models and reasoning tasks. It serves industries ranging from traditional AI to finance and science.

Versions and family

This generation follows the Trillium TPU v6e announced previously. Ironwood is also referred to as TPU v7 or TPU v7p in reports. It improves power efficiency by a factor of two compared to Trillium. Memory bandwidth increased by 4.5 times over the prior version. The chip replaces older formats like INT8 and BF16 with native FP8 support. No successor has been announced yet, as this product recently launched.

Why it matters

The chip aims to redefine infrastructure for the age of inference. Google claims a single pod outperforms the world’s largest supercomputer in specific metrics. It offers nearly thirty times better power efficiency than the first Cloud TPU from 2018. This efficiency helps address power constraints in data centres. By integrating hardware and software, Google seeks to lower costs for customers. The release reinforces Google’s position as a major provider of AI infrastructure alongside competitors.

Questions people ask about Ironwood

What is the primary purpose of the Ironwood chip?

Ironwood is designed specifically for AI inference workloads. It handles complex tasks like large language model reasoning and mixture of experts architectures. It also supports training but focuses on the computational demands of generating insights.

How much memory does each Ironwood chip contain?

Each chip provides 192 gigabytes of high bandwidth memory. This is six times the capacity of the previous Trillium generation. The increased memory helps process larger models without frequent data transfers.

What new calculation format does Ironwood support?

It is the first TPU to natively support FP8 calculations in its tensor cores. Previous generations used INT8 for inference and BF16 for training. This change allows for more efficient mathematical processing in AI models.

How is the Ironwood chip cooled in data centres?

The system uses advanced liquid cooling solutions. Google states this method reliably sustains twice the performance of standard air cooling. It is necessary to manage the heat generated by continuous heavy AI workloads.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on Ironwood yet, so the facts on this page come from the sources listed above: blog.google/products/google-cloud/ironwood-tpu-age-of-inference/, nextplatform.com/compute/2025/04/09/with-ironwood-tpu-google-pushes-th. 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

    The weekly digest

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