TPU
Google's custom-built processor for tensor operations, competitive with GPUs for training and serving.
Think of it like
An even more specialized worker designed just for neural network math.
Example
A TPU v4 pod has 256 TPU chips connected with ultra-fast networking, used to train Gemini.
How it actually works
TPUs are purpose-built for the exact operations neural networks need. They're very fast and power-efficient. The downside: they're only available on Google Cloud, less software support than GPUs, and harder to debug. Good for training large models, fine-tuning is usually cheaper on GPUs.
For product teams
Cost-effective for training at Google scale. Harder to use for serving individual requests.
For engineers
Consider TPUs if you're training a really large model. Expect learning curve.
Related
- Alternative to GPUs.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome