Google TPU v3: The Complete Guide

Google Cloud Next ’18 (24–26 July 2018) confirmed “Cloud TPU Pods and TPU v3 are now available in alpha” 2. The exact figure Google gave live on stage at I/O for peak pod performance is not independently confirmed in an official Google blog post; this page uses Google’s current documented spec instead of the press-reported keynote number.

Google’s own words: “Google first used liquid cooling in TPU v3 that was deployed in 2018,” and that switch let Google “double chip density” and build a liquid-cooled supercomputer “four times” the size of its air-cooled TPU v2 predecessor 3. This is confirmed at two separate official Google Cloud sources 13.

Per Google’s own current documentation 4: 32 GiB of HBM2 per chip at 900 GB/s, a 2D torus interconnect topology, 340 TB/s of pod-wide all-reduce bandwidth and 6.4 TB/s of bisection bandwidth. Google does not publish a process node for TPU v3.

Figures per Google Cloud’s current TPU v3 documentation 4. The 126-petaflops pod figure is Google’s current documented spec, which may differ from whatever number was stated live at the 2018 keynote; no official Google blog post confirming an exact keynote figure was found.

As with TPU v1 and v2, no official Google or Broadcom material ties a specific manufacturer to TPU v3. Google’s confirmed Broadcom manufacturing relationship is documented only from later generations.

Google’s own case studies from this generation include Recursion Pharmaceuticals, which cut a 24-hour local-GPU training run down to 15 minutes using TPU v3 Pods 6, and Google Translate, which used TPU v3 for both training and serving to push new models to production within hours of validation 7.

Checked September 2026, from Google Cloud’s own pricing page 5: TPU v3 pod allocations at $2.00 per chip-hour, standalone devices at $2.20 per chip-hour in europe-west4. These are legacy rates, still available nearly eight years after launch.

Google’s own November 2019 post states 32 Cloud TPU v3 devices processed 1 million ResNet-50 images per second in MLPerf Inference v0.5, with near-linear scaling from 1 to 32 devices, and calculated that processing all 7.7 billion of the world’s photographs at that rate would cost under $600 and take under 2.5 hours 7. A separate July 2019 post reports TPU v3 Pods running over 84% faster than the fastest on-premises systems measured in MLPerf Training 6. Both are independently audited MLPerf results, not purely company-stated comparisons.

TPU v3 succeeded TPU v2 primarily on raw compute and cooling capacity rather than a new capability like training. The next generation, TPU v4, introduced a genuinely new architectural feature: optical circuit switches that let Google rewire the interconnect between chips on the fly 1.

Google’s own inference-records post states Google Translate could push new models into production “within hours” of validation once TPU v3 handled both training and serving 7 — a workflow change that traces directly back to the liquid-cooling breakthrough that let this generation run hot enough to be worth building at all.