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
TPU v2 is where Google’s AI chip learned to do something its predecessor couldn’t: train new models, not just run finished ones. Announced 17–18 May 2017 1.
Deep diveThe technical detail
Google demonstrated a 64-chip pod delivering up to 11.5 petaflops at launch 1; Google Cloud documentation today lists a maximum slice size of 512 chips 2 — the pod size grew over time as Google expanded the product.
2.Launch and history
SimpleStart here
TPU v1 could only run models Google had already trained elsewhere on GPUs. TPU v2 closed that gap — the same chip family could now both teach a model and run it, and for the first time Google let outsiders rent the hardware.
Deep diveThe technical detail
Google’s own example from the launch announcement: a machine-translation model that took a full day to train on 32 of the best commercial GPUs available at the time trained in an afternoon on just one-eighth of a TPU v2 pod 1. The same announcement launched the TensorFlow Research Cloud — 1,000 Cloud TPUs given free to researchers who agreed to publish their results openly 3.
3.What’s inside it
SimpleStart here
TPU v2’s defining change from v1 is training support — the ability to run backward-propagation, not just forward inference, directly on the chip 1.
Deep diveThe technical detail
Each chip carries 64GB of high-bandwidth memory, a sharp jump from v1’s plain 8GB DRAM 1. Google does not publish a process node for TPU v2 in any official source found.
4.Full spec table
SimpleStart here
Up to 180 teraflops per chip and 64GB of memory; a 64-chip pod at launch delivered 11.5 petaflops 1.
Deep diveThe technical detail
| Spec | Google TPU v2 |
|---|
| Peak compute | Up to 180 teraflops per chip |
|---|
| Memory | 64GB high-bandwidth memory per chip |
|---|
| Launch pod (2017) | 64 chips, 11.5 petaflops |
|---|
| Current max slice | 512 chips |
|---|
| Process node | Not published by Google |
|---|
Launch figures per Google’s own 2017 announcement 1; current maximum slice size per Google Cloud’s TPU v2 documentation 2. A separate Google Cloud retrospective describes an intermediate “first TPU Pod” configuration of 256 chips 4 — pod size grew across several stages as Google scaled the product, and this page reports each figure with its own source rather than picking one.
5.Where it’s made
SimpleStart here
Google does not disclose who manufactured TPU v2 or on what process node. No official source names a foundry for this generation.
Deep diveThe technical detail
As with TPU v1, Google’s documented manufacturing partnership with Broadcom is confirmed only from later generations onward. No official Google or Broadcom material ties either company to TPU v2’s manufacturing by name.
6.Which systems use it
SimpleStart here
TPU v2 was the first TPU generation Google Cloud rented out to the public, alongside the free TensorFlow Research Cloud program for academic researchers 13.
Deep diveThe technical detail
It remains available today as a legacy Cloud TPU option, still listed on Google Cloud’s current pricing page nearly a decade after launch 5.
7.Official pricing
SimpleStart here
Google Cloud’s current pricing page lists TPU v2 as a legacy tier: $1.50/chip-hour for a pod allocation, or $1.2375–$1.305/chip-hour for a standalone device depending on region 5.
Deep diveThe technical detail
Checked September 2026, from Google Cloud’s own pricing page 5: TPU v2 pod allocations price at $1.50 per chip-hour in us-central1 and europe-west4; standalone TPU v2 devices price at $1.2375 to $1.305 per chip-hour depending on region. These are legacy rates — TPU v2 is no longer Google’s recommended option for new workloads, but remains rentable.
9.What came before, what came next
SimpleStart here
Came before: TPU v1 (inference only). Came after: TPU v3, announced 2018, which added liquid cooling to Google’s data centers for the first time.
Deep diveThe technical detail
TPU v2 succeeded TPU v1 by adding training capability 1. Google Cloud’s own 10-year TPU retrospective places TPU v3 as the next major generation, marked by the introduction of liquid cooling 4.
10.Hidden in plain sight
SimpleStart here
Google gave away 1,000 of these chips for free — the same year it started charging everyone else to rent them.
Deep diveThe technical detail
The exact same May 2017 announcement that introduced Cloud TPU as a paid rental service also launched the TensorFlow Research Cloud, giving 1,000 TPU v2 devices to academic researchers at no cost, on the condition that they publish their research openly 3 — a commercial launch and a giveaway program announced side by side.