Skip to content
Google's Ironwood TPU Is Rated at 4,614 Teraflops and 192 GiB of Memory per Chip, With 9,216 Chips Wired Into One Pod and General Availability on Google Cloud Since 31 March 2026
AI & Technology

Google's Ironwood TPU Is Rated at 4,614 Teraflops and 192 GiB of Memory per Chip, With 9,216 Chips Wired Into One Pod and General Availability on Google Cloud Since 31 March 2026

Illustration by tuput

English

Google's documentation lists each Ironwood chip at a peak of 4,614 teraflops in 8-bit arithmetic, with 192 GiB of memory and 7,380 GBps of memory bandwidth. Google's release notes record preview on 24 November 2025 and general availability on 31 March 2026.

· · 10 min read

Google’s documentation rates each Ironwood chip at a peak of 4,614 teraflops (trillions of arithmetic operations per second) in 8-bit arithmetic, with 192 GiB of high-bandwidth memory beside it, and lets up to 9,216 of those chips be linked into one pod. Ironwood is the seventh generation of the Tensor Processing Unit, or TPU, Google’s chip for the matrix arithmetic that AI models run on. Google’s release notes record it as generally available on Google Cloud from 31 March 2026.

Every figure here is Google’s own specification or peak rating, read from its pages as of 12 November 2026. No independent benchmark of Ironwood is cited, and a section near the end separates out the comparisons that rest on Google’s choice of test.

The numbers on Google’s specification page

Google’s TPU7x page (TPU7x is the cloud name for the first Ironwood release) sets the chip beside the two generations before it.

Per chip unless statedTPU v5pTPU v6e (Trillium)TPU7x (Ironwood)
Chips per pod8,9602569,216
Peak compute, BF16 (teraflops)4599182,307
Peak compute, FP8 (teraflops)4599184,614
Memory (GiB)9532192
Memory bandwidth (GBps)2,7651,6387,380
Chip-to-chip bandwidth, both directions (GBps)1,2008001,200

Google’s April 2025 announcement says the FP8 figures are emulated on v4 and v5p and native on Ironwood, which is why the v5p column repeats its BF16 number; the page does the same for v6e. Each Ironwood chip carries two TensorCores, the units that do matrix multiplication, and four SparseCores, which the April post describes as accelerators for the very large embeddings (lookup tables) used in ranking and recommendation. Four chips sit on each host computer, and the page lists 224 virtual CPUs and 960 GB of RAM for that four-chip machine.

What a teraflop is, and why the number format matters

A FLOP is one floating-point operation: one addition or multiplication on a number that can have a decimal part. A teraflop is a trillion of them per second and an exaflop is a billion billion. The 4,614 is a peak, the rate if every arithmetic unit is busy every moment, and the documentation labels it that way.

The label also names a number format. FP8 stores each number in 8 bits and BF16 (bfloat16) in 16. The documentation lists 2,307 teraflops for BF16 and exactly double for FP8, and the Google DeepMind textbook cited below says the chip does FP8 matrix multiplication at twice the BF16 rate. Fewer bits mean coarser numbers and more operations per second.

Multiplying out the pod figure: 9,216 chips at 4,614 teraflops is 42.5 exaflops, the number in Google’s April 2025 post and on Google’s Hot Chips 2025 slide, where it is labelled FP8. The same sum at the BF16 rating gives about 21 exaflops, which is the figure the textbook gives for a pod.

Why memory matters as much as arithmetic

The textbook, written by Google DeepMind engineers, describes a TPU as a core that specialises in matrix multiplication attached to a stack of fast memory called HBM, high-bandwidth memory. Bandwidth is how many bytes per second can move from that memory to the arithmetic units. Between the two sits a small on-chip scratchpad, which the textbook gives as 128 MiB on TPU7x, far faster than HBM but far smaller than its 192 GiB.

A chatbot writes its reply one token (a word or piece of a word) at a time, as tuput’s piece on how a chatbot guesses the next word describes, and the textbook notes that each step has to load all of the model’s parameters from HBM. The time to load them is therefore a floor under the time per step. Its worked example takes a 200-billion-parameter model in BF16, which is 400 GB, spread over 32 v6e chips: 12.5 GB per chip, read at about 1.6 TB per second, or roughly 7.8 milliseconds per step at best.

Run the same sum with Ironwood’s documented 7,380 GBps and 12.5 GB takes about 1.7 milliseconds. That is tuput’s arithmetic on a documented figure, for one model size over 32 chips, and a lower bound rather than a prediction of how fast any service replies. Google’s documentation itself says that despite its size, HBM can still be a bottleneck for some operations. Google’s pages round the bandwidth three ways: 7.37 TB/s in the April post, 7,380 GBps in the documentation and 7.3 TB/s on the Hot Chips slide.

A pod, and why its chips are wired as a torus

Google’s November 2025 post describes a pod as chips connected to one another so that they work as a single unit. For Ironwood the post gives up to 9,216 chips, joined by networking it rates at 9.6 terabits per second and sharing 1.77 petabytes of HBM. The documentation builds that from 144 cubes of 4x4x4 chips, which is 64 chips a cube, and a four-chip slice is the smallest configuration it lists. For slices larger than 2,048 chips the page says to contact Google’s account team.

Inside a cube each chip has direct links to its nearest neighbours in three directions. The page gives six links per chip at 200 GBps of two-way bandwidth each, 400 GBps along each axis and 1,200 GBps in total. For slices of 64 chips or more, the cubes are joined in a 3D torus, a grid whose edges wrap round to meet the opposite edge. Picture a row of chips: joining the last to the first turns it into a ring, and the textbook says this wraparound cuts the greatest distance between two chips from N to N/2. The textbook adds that slices smaller than a full cube get no wraparound, which it says typically doubles the time of most communication. It contrasts the design with GPU clusters, which connect chips through a hierarchy of switches.

Optical circuit switches join the cubes and can be reconfigured, so one pool of cubes can be cut into slices of different shapes. ServeTheHome’s report of the Hot Chips talk says a failed node can be dropped and the slice rebuilt on other racks from a saved checkpoint, and that 9,216 is a little above a power of two so that spare racks remain (8,192 would leave none). The April post puts a full pod at nearly 10 MW and says the chips are liquid cooled. How much power AI data centres draw in total is covered in tuput’s piece on IEA and Berkeley Lab forecasts.

Two chiplets under one lid

Google’s Hot Chips slides call Ironwood the first dual compute die TPU, and ServeTheHome reports the design lets Google build beyond a single die. Each chiplet has one TensorCore, two SparseCores and 96 GiB of HBM, according to the documentation, and a die-to-die link six times faster than one chip-to-chip link joins them. Software sees each chip as two devices: on a four-chip machine, the JAX library lists eight.

The page says TPU7x runs the JAX and PyTorch frameworks and does not support TensorFlow. For serving models it lists the vLLM and SGLang inference engines. Google’s 9 November 2026 post says it open-sourced TorchTPU the month before, so that PyTorch code and Hugging Face models run on TPUs without rewrites.

Dates, zones and customers

Google announced Ironwood on 9 April 2025 at Cloud Next 25, calling it “the first designed specifically for inference”. Its documentation and November 2025 post describe it as built for training and inference both. Google’s slides for the Hot Chips conference on 26 August 2025 name Norman Jouppi and Sridhar Lakshmanamurthy as presenters and say deployment at hyperscale was under way.

On 23 October 2025 Anthropic announced plans to use “up to one million TPUs”, an expansion it called worth tens of billions of dollars and expected to bring well over a gigawatt of capacity online in 2026. That post names no TPU generation. Google’s 6 November 2025 post places the same plan in its Ironwood section and quotes Anthropic’s head of compute, James Bradbury, on Ironwood’s inference and training gains. It also quotes Lightricks and Essential AI as early users.

The November post’s text said Ironwood would be generally available “in the coming weeks”. Google’s release notes carry the dates: preview on 24 November 2025 and general availability on 31 March 2026. The TPU7x page does not list locations. The separate regions and zones page lists two zones for TPU7x as of 12 November 2026, us-central1-ai1a and us-central1-c, both in Google’s us-central1 region.

For India, the PIB backgrounder of 10 February 2026 says the IndiaAI Mission had onboarded more than 38,000 GPUs at Rs 65 per hour and also 1,050 TPUs, which it defines as chips built specifically for AI tasks. It does not name a TPU generation, so nothing in it says whether any are Ironwood. The mission is covered in tuput’s IndiaAI Mission article, and the national petaflop machines in its piece on the PARAM supercomputers.

Which comparisons are Google’s own

Google’s November 2025 post says Ironwood offers a 10 times peak performance improvement over TPU v5p and more than 4 times better performance per chip for training and inference than v6e. The Hot Chips slide also says more than 10 times the compute of v5p. The table shows how much rides on the number format. In FP8, 4,614 against v5p’s 459 is 10 times, but v5p’s column is emulated FP8, and at BF16 the ratio is 5 times (2,307 against 459). Against v6e, on the page’s figures, it is 5 times at FP8 and 2.5 times at BF16. The post does not say which workloads or number format sit behind its “more than 4 times”.

The memory ratios in the April post follow from the table and are specifications, not test results: 192 GiB against 32 is 6 times, and 7,380 against 1,638 GBps is 4.5 times. The same post claims twice the performance per watt of Trillium and nearly 30 times the efficiency of Google’s first Cloud TPU of 2018, with a chart note saying the measure is peak FP8 flops per watt of thermal design power. ServeTheHome wrote of the same 2 times claim at Hot Chips that it was unclear whether it was measured at the same number format.

The April post also sets the 42.5-exaflop pod against El Capitan, which it lists at 1.7 exaflops. Lawrence Livermore National Laboratory reports that machine at 1.742 exaflops on the High-Performance Linpack test in the June 2025 TOP500 list, and at 16.7 exaflops on a separate mixed-precision test. One figure is a theoretical peak at FP8 and the other a measured benchmark, so they answer different questions.

After Ironwood

Google’s 9 November 2026 post, by chief technologist Amin Vahdat, says eight generations of TPUs now exist and that two eighth-generation chips, 8t for training and 8i for reasoning, were introduced at Cloud Next in April 2026. It quotes up to 2.7 times better training performance and 80 percent better inference price performance than the previous generation, both Google’s figures. The same post adds TPU 8tx, for confidential inference, and TPU 8i Max, which puts two 8i chips on one package with double the compute, SRAM and memory bandwidth of the 8i, and says the company has introduced four TPU systems in one year. It describes the TPU business as having nearly 100 customers. It gives no chip-level specification for any of the four, so none are compared with Ironwood here.

Share
Copied!

Sources & further reading

  1. Google Cloud documentation: TPU7x (Ironwood), last updated 11 November 2026
  2. Google Cloud documentation: TPU regions and zones
  3. Google Cloud documentation: Cloud TPU release notes
  4. Google: Ironwood, the first Google TPU for the age of inference (9 April 2025)
  5. Google Cloud blog: Ironwood TPUs and new Axion-based VMs for your AI workloads (6 November 2025)
  6. Google Cloud blog: Our newest TPUs, 8tx and 8i Max, built for the agentic era (9 November 2026)
  7. Hot Chips 2025: Ironwood, Delivering Best in Class perf, perf/TCO and perf/Watt for Reasoning Model Training and Serving (Google slides, 26 August 2025)
  8. Anthropic: Expanding our use of Google Cloud TPUs and Services (23 October 2025)
  9. Google DeepMind authors: How to Scale Your Model, Part 2, How to Think About TPUs
  10. ServeTheHome: Google Ironwood TPU Swings for Reasoning Model Leadership at Hot Chips 2025 (26 August 2025)
  11. Lawrence Livermore National Laboratory: El Capitan reigns supreme across three major supercomputing benchmarks (16 June 2025)
  12. PIB Backgrounder: Democratising AI in India (10 February 2026)

Researched and written with the help of AI tools and edited for accuracy. Provided for general information and discussion only, not professional advice. See our editorial standards and disclaimer. Spotted an error? Tell us.

#google ironwood#tpu#ai accelerator#tpu7x#high bandwidth memory#ai inference#ai hardware#google cloud

Enjoyed this? Get the next one.

One good read at a time, straight to your inbox. No spam, unsubscribe anytime.

More in AI & Technology
Eight Google Authors Posted the Transformer Paper on 12 June 2017, and Its Big Model Scored 28.4 BLEU in English to German After 3.5 Days on Eight GPUs
A walk through the design behind today's chatbots, using the paper's own tables and Google's own example of a word that has to look at a river. It also covers the parts of the 2017 design that later models kept, changed or dropped.
Bharat-VISTAAR, Launched in Jaipur on 17 February 2026, Lets a Farmer Ask an AI Adviser Questions by Phone Call on 155261 and Draws on ICAR, IMD, Agmarknet and AgriStack Data
One call to 155261 reaches an AI adviser fed by ICAR practice guides, weather and mandi data. The launch releases, the language plan and the usage figures published so far, set out with dates and sources.
Ramayana Part 1 Reached English-Speaking Cinemas on 6 November 2026 With Human Voice Performances and Brahma AI Matching the Actors' Lips, According to Its Makers
A two-step process, human performance first and software second, is how the producers describe the film's English release. The Television Academy, Brahma AI's own pages and several reviewers fill in the rest.
← all articles