AI Chips · Product page

Google TPU v1: The Complete Guide

Blueprint to production floor in 15 months. It ran for a year before the outside world knew it existed. Simple to expert, every number links to its source.

Ask an AI about this chip:CAGPX
Google TPU v1 10 sections · 2 levels 3 linked sources Checked September 2026
How to read this page. Each section starts Simple, then goes to a Deep dive: stop wherever you have what you need. The small numbers are sources: click one to open the original document. Where the maker has not published a figure -- a price, a die size, a factory address -- this page says so rather than estimate it.

1.At a glance

SimpleStart here

The TPU v1 was Google’s first custom AI chip, publicly announced 19 May 2016. Google said it had already been running in its own data centers “for more than a year” 1 — powering search results, Street View, and a Go-playing program the world was about to hear a lot about.

Deep diveThe technical detail

Google’s own paper states the chip went from design to production deployment in just 15 months 2 — and that it had already been running quietly inside Google’s own data centers for a year before the public announcement 1.

2.Launch and history

SimpleStart here

TPU v1 only ever ran already-trained models — it could not train new ones. Google built it purely to make its own products faster and cheaper to run at scale, not to sell or rent to anyone.

Deep diveThe technical detail

Google’s 2016 announcement named three products the chip was already powering: RankBrain (search ranking), Street View, and “the matches against Go world champion Lee Sedol” 1 — a reference to AlphaGo’s landmark March 2016 victory. Google framed the chip’s efficiency gain as equivalent to “fast-forward[ing] technology about seven years into the future” 1.

3.What’s inside it

SimpleStart here

An 8-bit matrix-multiplication engine built for one job: running already-trained neural networks fast and cheap. No training capability, no HBM — just plain DRAM and a PCIe connection 2.

Deep diveThe technical detail

65,536 8-bit multiply-accumulate units, clocked at 700 MHz, on a die no larger than half the size of a contemporary Intel Haswell server chip 2. 28 MiB of on-chip memory (24 MiB unified buffer, 4 MiB accumulators, 64 KiB weight FIFO), connected over PCIe Gen3 x16 as an add-in card rather than a networked system — TPU v1 had no chip-to-chip interconnect or pod concept at all 2.

4.Full spec table

SimpleStart here

92 TeraOps/second of 8-bit compute, 8GB of plain DRAM (not HBM) at 34 GB/s, and a 40W power draw — a fraction of what a contemporary GPU used 2.

Deep diveThe technical detail
SpecGoogle TPU v1
Peak compute92 TeraOps/second (8-bit)
Clock speed700 MHz
Process node28 nm
On-chip memory28 MiB
Off-chip memory8 GiB DRAM (not HBM), 34 GB/s bandwidth
Power40W busy / 28W idle (chip); 384W busy (full 4-chip host system)
Host interfacePCIe Gen3 x16
Multi-chip interconnectNone — single-card accelerator

Figures per Google’s own ISCA 2017 paper 2. Process node is the only Google TPU generation for which Google has published the node size directly.

5.Where it’s made

SimpleStart here

Google does not state who manufactured TPU v1 or on what process, beyond the 28-nanometer node disclosed in its own paper. No foundry is named in any official Google source.

Deep diveThe technical detail

Google’s long-running manufacturing partnership with Broadcom for later TPU generations is well documented from TPU v5 onward, but no official Google or Broadcom source found ties Broadcom, or any other specific manufacturer, to TPU v1 by name. This page does not assert one.

6.Which systems use it

SimpleStart here

TPU v1 was never offered for rent. It ran only inside Google’s own data centers, serving Google’s own products.

Deep diveThe technical detail

Confirmed by Google Cloud’s own current pricing page, which lists no TPU v1 tier at all 3 — unlike TPU v2 and later generations, which remain listed as legacy rentable options. TPU v1 predates Cloud TPU as a product entirely.

7.Official pricing

SimpleStart here

There was never a price. TPU v1 was internal-only and never appeared on any Google Cloud price list 3.

Deep diveThe technical detail

Google Cloud’s pricing page lists legacy per-chip-hour rates going back to TPU v2, but no TPU v1 entry exists at any point in that history 3 — consistent with TPU v1 never being offered outside Google’s own infrastructure.

8.Real-world performance

SimpleStart here

Google’s own paper reports TPU v1 ran 15 to 30 times faster than the GPUs and CPUs of its era, at 30 to 80 times better performance per watt 2 — company-published figures, not an independent benchmark.

Deep diveThe technical detail

The paper compares TPU v1 against an NVIDIA K80 GPU and an Intel Haswell CPU of the same era, both in wide production use at the time 2. A hypothetical variant using faster GDDR5 memory instead of plain DRAM was modeled (not built) at up to 70x the speed and 200x the performance-per-watt 2.

9.What came before, what came next

SimpleStart here

Came before: nothing — TPU v1 was Google’s first custom AI chip. Came after: TPU v2, announced May 2017, which added the ability to train models, not just run them.

Deep diveThe technical detail

TPU v1 has no predecessor in Google’s chip lineup. Its single limitation — inference only, no training — became the headline feature of TPU v2 a year later 2.

10.Hidden in plain sight

SimpleStart here

The chip that helped beat the world Go champion was announced to the public only after the match was already won.

Deep diveThe technical detail

Google’s own March 2016 blog post about AlphaGo’s victory over Lee Sedol makes no mention of TPUs at all — the chip wasn’t revealed publicly until two months later, in the 19 May 2016 announcement that first named it as having powered those matches 1. For two months, the hardware behind one of AI’s most-watched moments was simply unknown to the outside world.

11.Sources

3 sources, checked September 2026. Where NVIDIA or another maker has not published a figure, this page says so rather than estimate it.

  1. Google Cloud Blog: Google Supercharges Machine Learning Tasks with TPU Custom ChipOfficial
  2. Jouppi et al. (Google), ISCA 2017: In-Datacenter Performance Analysis of a Tensor Processing UnitPaper
  3. Google Cloud: Cloud TPU PricingOfficial