TPU
Google's custom Tensor Processing Unit, an ASIC designed specifically for accelerating machine learning workloads.
Technical explanation
TPU stands for Tensor Processing Unit, and it's Google's answer to the GPU bottleneck in AI training. Unlike GPUs, which were repurposed from graphics work, TPUs were built from scratch specifically for machine learning. They're application-specific integrated circuits (ASICs) optimized for tensor operations, the mathematical foundation of neural networks.
Google introduced TPUs back in 2016 and has been iterating ever since. The latest versions, like TPU v5p, can train models faster and more efficiently than many GPU alternatives. They're especially good at Google's own TensorFlow framework, though JAX support has gotten really strong too. If you're using Google Cloud's Vertex AI, you'll likely run into TPUs as a compute option.
The trade-off with TPUs is flexibility. They're locked into Google's ecosystem. You can't buy TPU hardware and run it in your own data center. You access them through Google Cloud, period. For some teams, that's a dealbreaker. For others, the performance gains and tight integration with Google's ML stack make it worthwhile.
Cost-wise, TPUs can be competitive with high-end GPUs, especially for large training jobs. Google offers TPU pods that scale massively for enterprise workloads. If you're building on TensorFlow or JAX and already committed to Google Cloud, TPUs deserve serious consideration. Otherwise, NVIDIA GPUs remain the more flexible choice.