TPU
We explain Google's in-house AI chip: the TPU roadmap from Trillium to new generations, and its real role in training large models and serving them from Google Cloud.
The TPU is Google's purpose-built AI chip: launched in 2016 and iterated through Trillium to newer architectures, it trains models like Gemini and serves compute via Google Cloud. Deeply optimized for matrix math, it underpins Google's AI infrastructure — and Project Suncatcher will test it in orbit. We track its architectural evolution, measured performance, cloud availability, and how it really stacks up against NVIDIA GPUs.
