Google's Tensor Processing Unit (TPU) first deployed internally in 2015 as a proprietary AI accelerator for cloud-based machine learning workloads in data centers. Unlike GPUs, TPUs use a systolic array architecture — a 2D grid of multipliers that passes calculation outputs directly to the next unit without writing back to memory, eliminating bottlenecks for massive large language model (LLM) training. This makes TPUs more power-efficient for companies like Anthropic and Midjourney serving billions of daily AI requests.

The term TPU recently appeared in consumer devices with the Pixel 11 smartphone's Google Tensor G6 chip, claiming 50 percent more TPU compute. However, this on-device TPU essentially functions as a Neural Processing Unit (NPU), handling camera processing and local AI tasks. Google claims up to 3.5 times faster AI processing while using up to 3.5 times less energy versus previous generations.

GPUs remain versatile jack-of-all-trades chips for gaming, 3D rendering, AI training, and crypto mining. NPUs in modern smartphones, Macs, and PCs handle on-device generative AI like photo editing and ChatGPT tasks. The key distinctions lie in scale and use case: TPUs dominate cloud-scale AI, NPUs and on-device TPUs serve consumer devices, while GPUs bridge both worlds with broader flexibility.