Skip to content
Working Knowledge
GuideReferenceWorkbenchBoardNotesAbout
Search⌘K

Browse categories

The Reference

01The Big Picture02Model Architectures03How Models Learn04How Models Run05Getting Models to Act06Language & Perception07Measuring Models08Safety & Alignment09Hardware & Infrastructure10Version Control11The Command Line12Databases & SQL13Environments & Packaging14Shipping & Collaboration15The Workbench
09 · Hardware & Infrastructure

TPU

Tensor Processing Unit

InterInvestors

Google's custom in-house AI chip, designed specifically for neural networks. Its main lever to avoid total dependence on Nvidia.

Concrete example

Google trains and serves much of its own AI on TPUs rather than buying Nvidia for everything.

Nearby in Hardware & Infrastructure

NVLink / InfiniBandHigh-speed GPU interconnectsTCOTotal Cost of OwnershipVRAMVideo RAM
Previous term← TPSNext termTraining Data →

AtlasScope

I explain technical systems in plain English—and build practical tools for using them.

Read the guideBrowse the referenceWhat Atlas is doing now

Contact & collaboration

GitHub ↗
Working Knowledge · A practical field guideBuilt and maintained by Atlas
HomeGuideSearchWorkbenchMore