Decoding TPUs: Google's Custom AI Chip vs Traditional GPUs

Google's Tensor Processing Unit (TPU) is a proprietary AI accelerator originally built for data center workloads, but the Pixel 11's Tensor G6 chip uses a TPU to handle on-device AI tasks, replacing the typical NPU. Unlike GPUs, which are general-purpose parallel processors, TPUs are optimized for the specific mathematical operations used in machine learning models. Google claims the Pixel 11's TPU delivers significantly faster AI processing with reduced power consumption compared to earlier NPUs.

Google's TPU architecture relies on a systolic array—a 2D grid of multipliers where each unit's output feeds directly into the next, avoiding constant memory write-backs. This differs from GPUs, which shuttle data between compute cores and high-bandwidth memory, creating potential bottlenecks. The design suits massive deep-learning workloads like those powering Anthropic and Midjourney's daily billions of requests.

In the Pixel 11's Tensor G6, Google applies the TPU name to a smaller on-device accelerator handling camera processing and local AI tasks. The company positions this as a replacement for the NPU found in rival phones, claiming substantial gains in processing speed and energy efficiency over earlier neural processors.