Get the newsletter
Google · Chip · Filed Feb 6, 2026 · USPTO published Aug 13, 2026 · Aug 13 release

Google Patents an AI System That Adjusts Chip Power Based on What You're Doing

Every chip inside your phone is constantly doing more work than it needs to, because it can't tell the difference between a quick email check and a graphics-heavy game. Google's new patent describes an AI layer that watches what the chip is actually doing and dials power up or down accordingly.

System-on-chip architecture linking internal processing blocks with memory and external user devices. Drawing from patent filing US 2026/0236083 A1.
System-on-chip architecture linking internal processing blocks with memory and external user devices.All 6 drawings ↓
Publication number
US 2026/0236083 A1(US20260236083A1)
Applicant
Google LLC
Filed · Published
Feb 6, 2026 · Aug 13, 2026
Inventors
Shen-Hong Wang, Yu-Jin Yang, Hee jun Park
Status when we published
Waiting for an examiner
Document
20 claims
US classification
713/320
Parent application
Claims priority from a provisional application 63758190 (filed Feb 13, 2025)

How Google's AI power manager reads your chip's workload

Every time you tap your phone awake, a handful of chips start running at a fixed gear even if all you need is to glance at a notification. They don't inherently know whether you're about to play a demanding 3D game or just scroll through a grocery list, so they tend to over-prepare and burn more battery than necessary.

Google's patent describes a power manager that uses a small AI model to watch built-in chip counters, the tiny scorekeepers that track how hard each part of the processor is working, and then classify what kind of task is actually running. Once it labels the task, it looks up a matching "power hint" in a table and adjusts the processor's speed and energy settings to fit.

The result, in theory, is a chip that wastes less power on light tasks and still has headroom for heavy ones. You'd feel this as longer battery life without any slowdown during the moments that actually need speed.

From the filing · CLAIM 1
… determining, by an ML classifier model of the power manager, a workload type in response to processing the plurality of inputs based on a clustering algorithm applied to the plurality of inputs by the ML classifier model …

Translation: The system uses machine learning to categorize exactly what kind of task your device is currently performing.

How the ML classifier maps workloads to power settings

The system centers on a power manager that sits alongside the processor inside a system-on-chip (SoC), the single integrated circuit that houses the CPU, GPU, and memory controller in modern phones and tablets.

Here's the sequence:

  • Hardware performance counters (built-in meters that track cache misses, memory bandwidth, instruction throughput, and similar metrics) feed a stream of numbers to the power manager.
  • An ML classifier model applies a clustering algorithm (a technique that groups similar data points together, like sorting emails by topic) to identify which of several predefined workload types the chip is currently handling, such as video decode, browsing, or gaming.
  • The identified workload type is used to look up a power hint in a mapping table, a pre-built chart that links each workload category to a recommended power-and-performance profile.
  • Those profile settings, covering things like clock frequency and power caps, are then applied to a specific processor block within the SoC.

The patent emphasizes "fine-grained" control, meaning the system can adjust individual processor blocks independently rather than scaling the whole chip up or down as a unit. One or more neural network models are listed as part of the power manager's design, suggesting the classifier can be more expressive than a simple lookup table alone.

From the filing · THE ABSTRACT
A power hint is identified based on the workload type and using a data structure that maps workload types to power hints. For the determined workload type, power and performance policy settings are determined for a processor block of the system-on-chip based on the identified power hint.

Translation: Once the task is identified, the system looks up a power setting to adjust how much energy the processor uses.

What this means for Android battery life and Tensor chips

Battery life is one of the most complained-about aspects of modern phones, and much of the waste happens not during intensive tasks but during the hundreds of in-between moments when chips run hotter than they need to. A workload-aware power manager that reacts in real time, rather than relying on static profiles set by engineers before shipping, could meaningfully close that gap without sacrificing the speed users notice.

For Google, this is directly relevant to its in-house Tensor chips inside Pixel phones, where tight integration between hardware and software is the main selling point. The approach also fits neatly into Android's broader power management framework. Power-efficiency patents like this one are among the newest Big Tech patents in the chip-software integration space, where Google, Apple, and Qualcomm are all competing to squeeze more from the same silicon.

Source: US 2026/0236083 A1 at the USPTO (PDF). Full patent text and figures are in the official publication.

Get our take in your Top Stories

Liked this breakdown? Add Patentlyze as a preferred source on Google, and our plain-English take shows up more often in your Top Stories the next time Google patent news breaks.

There are more where this came from

We read every patent application Big Tech publishes and send you the ones worth knowing. Plain English, free, every week.

The drawings · 6 sheets

6 drawing sheets from US 2026/0236083 A1 · click any drawing to enlarge

System architecture and connected devices, drawing sheet 1 from US 2026/0236083 A1
Sheet 1 System architecture and connected devices
Power management and control flow, drawing sheet 2 from US 2026/0236083 A1
Sheet 2 Power management and control flow
Workload classification and statistics, drawing sheet 3 from US 2026/0236083 A1
Sheet 3 Workload classification and statistics
Workload collection and categorization, drawing sheet 4 from US 2026/0236083 A1
Sheet 4 Workload collection and categorization
Reinforcement learning policy mapping, drawing sheet 5 from US 2026/0236083 A1
Sheet 5 Reinforcement learning policy mapping
Power adjustment process steps, drawing sheet 6 from US 2026/0236083 A1
Sheet 6 Power adjustment process steps
Patent filing page

Watchlists this filing belongs to

All watchlists →