Short answers, pulled from the story.
The Tensor Processing Unit is a custom silicon chip deployed by Google in 2015 to handle neural network tasks more efficiently than standard graphics cards. This bespoke creation was designed to solve the energy demands of running neural networks at a global scale.
The Tensor Processing Unit uses a systolic array design that allows data to flow rhythmically through the chip like blood cells in a vein instead of shuttling data back and forth between a processor and memory. This architecture eliminates the memory bottleneck by moving data through a grid of 256 by 256 multipliers in a synchronized dance.
Google opened the Tensor Processing Unit to third-party companies on the 12th of February 2018 through its cloud computing service. This move transformed the chip from an internal efficiency tool into a commercial product that challenges the dominance of Nvidia in the AI accelerator market.
The Edge TPU is a purpose-built ASIC that consumes only 2 watts of power and performs 4 trillion operations per second. This chip was integrated into the Pixel 4 smartphone as the Pixel Neural Core to handle camera features with minimal latency and power consumption.
Singular Computing filed a lawsuit against Google in 2019 alleging patent infringement regarding the dynamic range of floating-point numbers used in the chip. The dispute centered on the bfloat16 format, a non-standard floating-point format that Google invented to maximize the efficiency of its neural networks.
The Trillium chip announced in May 2024 offers a 4.7 times performance increase over the previous v5e model. The Ironwood chip announced in April 2025 promises a peak computational performance rate of 4,614 teraflops per second.