Get the newsletter
Google · Chip · Filed Aug 6, 2025 · USPTO published Aug 13, 2026 · Aug 13 release

Google Patents a Dedicated Chip Component to Speed Up AI Math Calculations

Google is patenting a small but specialized piece of chip architecture whose only job is to figure out which part of a processor handles which slice of an AI calculation, so nothing waits around for instructions that are stuck in the wrong place.

Three-dimensional cubes representing inputs, filters, and output data manipulated during artificial intelligence calculations. Drawing from patent filing US 2026/0236732 A1.
Three-dimensional cubes representing inputs, filters, and output data manipulated during artificial intelligence calculations.All 9 drawings ↓
Publication number
US 2026/0236732 A1(US20260236732A1)
Applicant
Google LLC
Filed · Published
Aug 6, 2025 · Aug 13, 2026
Inventors
Yang Yang, Suyog Gupta, Vishal Anand, Temitayo Fadelu, Xiaoxiao Long
Status when we published
Waiting for an examiner
Document
20 claims
US classification
706/27
Parent application
National Stage Entry of PCT/US2023/012817 (filed Feb 10, 2023)(PCTUS2023012817)

What Google's tensor mapping unit actually does inside a chip

Every time a Google data center runs an AI model, thousands of individual math operations have to get sorted and sent to the right part of a processor chip. That sorting step sounds simple, but at the speeds modern AI chips run, even a small delay in routing work to the right spot wastes real time and real energy.

Google's new patent describes a dedicated unit baked directly into the chip whose only job is that routing work. Instead of making the main processor figure out where each piece of a calculation lives, this unit reads an instruction, maps out exactly which positions in a data array belong to which slice of the chip, and fires off the precise signals needed to fetch those numbers from the right local memory. The main processor never has to stop and sort things out itself.

The practical effect is that AI math, particularly the kind of large matrix operations that power tools like Google's own language models, can move through the chip with less overhead and fewer stalls.

From the filing · CLAIM 1
An integrated circuit configured to accelerate tensor computations for a machine-learning model implemented at the integrated circuit, the integrated circuit comprising: a processing engine; and a tensor mapping unit configured to: generate a plurality of tensor element positions for a tensor based on an instruction received from the processing engine …

Translation: The chip includes a specialized unit that calculates exactly where data should go to speed up artificial intelligence tasks.

How the mapping unit assigns tensor positions to compute tiles

The patent covers a chip component called a tensor mapping unit (TMU). A tensor, in this context, is just a structured grid of numbers, the same kind of data structure that AI models constantly multiply, add, and transform. Chips built for AI work tend to break those grids into chunks and farm the chunks out to many small processor sections called compute tiles, each with its own local memory.

The problem the TMU solves is coordination. When the main processing engine issues an instruction, something has to determine which tile owns which chunk of the tensor and then fetch the right numbers from that tile's memory. Today that bookkeeping can sit with the central processor, adding overhead. The TMU offloads it entirely.

  • Position generation: The TMU reads the incoming instruction and works out the full list of element positions inside the tensor (think: row and column addresses in the data grid).
  • Tile assignment: It checks which positions belong to which compute tile on the chip.
  • Control signal dispatch: It sends precise fetch commands directly to the correct tile memory, bypassing the central processor for that work.

The result is a tighter pipeline: each compute tile gets its inputs faster because the routing decision is made by dedicated hardware rather than shared processing resources. The patent frames this as an acceleration mechanism specifically for machine-learning models running on custom integrated circuits, which points squarely at Google's in-house Tensor Processing Unit (TPU) line.

From the filing · THE ABSTRACT
… determine that a first subset of the tensor element positions is assigned to a first compute tile of the integrated circuit; and iii) generate control signals to obtain a group of inputs from a tile memory of the first compute tile based on the first subset.

Translation: The system identifies which specific part of the chip should handle a piece of data and then pulls that data from memory.

What this means for Google's custom AI chip strategy

Google designs its own AI chips, the TPU family, specifically to run the matrix math that neural networks demand. Any architecture that reduces the cycles the main processor spends on bookkeeping means more of the chip's budget goes toward actual computation. For data centers running AI inference at scale, that kind of efficiency compounds quickly across millions of requests.

The TMU concept is also a signal about how Google thinks about chip specialization: offload every repeatable coordination task to a fixed-function unit so the programmable cores stay free. That philosophy has driven the TPU roadmap for years, and this patent extends it one level deeper into the memory-fetch pipeline. Coverage of new Big Tech patents in the AI chip space shows this kind of fine-grained memory-routing work is where the real architectural competition is happening right now, well below the level of headline transistor counts.

Source: US 2026/0236732 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 the new patents from Big Tech, carmakers and more and send you the ones worth knowing. Plain English, free, every week.

The drawings · 9 sheets

9 drawing sheets from US 2026/0236732 A1 · click any drawing to enlarge

Processor architecture with memory blocks, drawing sheet 1 from US 2026/0236732 A1
Sheet 1 Processor architecture with memory blocks
Compute unit with multiple memory registers, drawing sheet 2 from US 2026/0236732 A1
Sheet 2 Compute unit with multiple memory registers
Request response processing pipeline, drawing sheet 3 from US 2026/0236732 A1
Sheet 3 Request response processing pipeline
Virtual address memory mapping layout, drawing sheet 4 from US 2026/0236732 A1
Sheet 4 Virtual address memory mapping layout
Physical tensor slicing layout, drawing sheet 5 from US 2026/0236732 A1
Sheet 5 Physical tensor slicing layout
Data item parameter configuration table, drawing sheet 6 from US 2026/0236732 A1
Sheet 6 Data item parameter configuration table
Continuous space versus physical memory layout, drawing sheet 7 from US 2026/0236732 A1
Sheet 7 Continuous space versus physical memory layout
Machine learning instruction execution flowchart, drawing sheet 8 from US 2026/0236732 A1
Sheet 8 Machine learning instruction execution flowchart
Show sheet 9
Google Patents a Dedicated Chip Component to Speed Up AI Math Calculations patent drawing, sheet 9
Sheet 9
Patent filing page

Watchlists this filing belongs to

All watchlists →