Get the newsletter
Groq · Chip · Filed Jul 27, 2026 · USPTO published Nov 19, 2026 · This week's release

Groq Files Patent for an AI Chip That Plans Every Data Move in Advance

Most chips spend effort working out what each piece of data is for. Groq's filing describes a chip where the software has already planned every move, so the data needs no labels at all.

An AI chip moves numbers step by step through storage and calculation areas to complete tasks on a fixed schedule. Drawing from patent filing US 2026/0368089 A1.
An AI chip moves numbers step by step through storage and calculation areas to complete tasks on a fixed schedule.All 17 drawings ↓
Publication number
US 2026/0368089 A1(US20260368089A1)
Applicant
Groq, Inc.
Filed · Published
Jul 27, 2026 · Nov 19, 2026
Inventors
Jonathan Alexander Ross, Dennis Charles Abts, John Thompson, Gregory M. Thorson
Status when we published
Examiner raised objections; the company can respond
Document
20 claims
US classification
712/220
Examiner
DUDEK JR, EDWARD J (Art Unit 2132)
Parent application
Continuation of 19/301,427 (filed Aug 15, 2025)(19301427)

What Groq's scheduled-by-software chip does

A train station can run without announcements, and nobody minds. Every train arrives at a fixed minute, so each platform crew already knows which train it is and what to do with it. Groq's new patent applies that idea to the inside of a computer chip.

Most chips attach labels to data and spend effort figuring out what each piece is for and where it should go. In Groq's design, a compiler (the software that turns a program into chip instructions) makes the whole schedule ahead of time. Data moves through the chip in set lanes, and each section knows what to do from when the data shows up.

The filing is aimed at AI work, where chips shuffle huge piles of numbers. Less guessing and less label-handling means fewer delays, which could mean faster answers from the AI services you use.

From the filing · CLAIM 1
… a streaming register file comprising a plurality of stream registers coupled between the at least one memory functional group and the at least one execution functional group to transfer the operand data and result data from execution of the operations on the operand data …

Translation: A special memory system hands data back and forth between storage areas and the math units.

How Groq's compiler keeps data from colliding

Groq's chip is split into slices, each built for one job: storing data, doing math, or switching data between lanes. Instructions travel along each slice in one direction, while data travels across the chip at a right angle to them. Because the compiler (the software that turns a program into chip instructions) knows the hardware exactly, it times everything so data and instructions meet at a planned moment.

The claims focus on one piece of this: a streaming register file (a set of small holding spots that sit between the memory and the math units and pass data back and forth). The compiler writes the instructions that say which holding spot is used and when. The method claims add that the compiler sets execution times so two math units never reach for the same spot at once.

The dependent claims add details:

  • Holding spots are arranged by stream number and lane, and can be grouped to carry larger number types.
  • The compiler keeps a clock-tick-accurate model of the chip to plan the data flow.
  • An unused bit in each spot flags errors, such as an out-of-range memory lookup or two accesses colliding in the same memory bank.
From the filing · THE ABSTRACT
… allowing operand data to be transmitted between the slices of the processor without any accompanying metadata. Instead, each slice is able to determine what operations to perform on received data based upon the timing at which the data is received …

Translation: Data travels without labels because chips know what to do based on exact arrival times.

Why Groq's pre-planned chip design matters

If you use an AI chatbot or image tool, a chip like this sits behind the screen, and time wasted inside it becomes time you spend waiting. A design that skips labeling and guessing could make answers arrive faster or let the same chip serve more people, though the filing gives no speed numbers.

The filing describes a processor aimed at machine learning work, and it covers a core piece of how that chip is organized. The examiner's first review rejected the claims as filed, which is routine and usually followed by amendments, so the final scope may end up narrower than what you see here.

Source: US 2026/0368089 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 Groq 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 · 17 sheets

17 drawing sheets from US 2026/0368089 A1 · click any drawing to enlarge

A grid of repeating cores on an AI chip, drawing sheet 1 from US 2026/0368089 A1
Sheet 1 A grid of repeating cores on an AI chip
Dividing the chip into specialized vertical columns, drawing sheet 2 from US 2026/0368089 A1
Sheet 2 Dividing the chip into specialized vertical columns
Sending program commands upward through each column, drawing sheet 3 from US 2026/0368089 A1
Sheet 3 Sending program commands upward through each column
Passing calculation data sideways across the chip, drawing sheet 4 from US 2026/0368089 A1
Sheet 4 Passing calculation data sideways across the chip
Routing numbers through math blocks and back into memory, drawing sheet 5 from US 2026/0368089 A1
Sheet 5 Routing numbers through math blocks and back into memory
Splitting onboard memory into two matching halves, drawing sheet 6 from US 2026/0368089 A1
Sheet 6 Splitting onboard memory into two matching halves
How memory units connect to internal communication lanes, drawing sheet 7 from US 2026/0368089 A1
Sheet 7 How memory units connect to internal communication lanes
Stacking data paths so math units work together, drawing sheet 8 from US 2026/0368089 A1
Sheet 8 Stacking data paths so math units work together
Show the other 9 sheets
Two-way traffic carries numbers across the processing lanes, drawing sheet 9 from US 2026/0368089 A1
Sheet 9 Two-way traffic carries numbers across the processing lanes
A single memory slot packs error checks alongside raw data, drawing sheet 10 from US 2026/0368089 A1
Sheet 10 A single memory slot packs error checks alongside raw data
How each column receives and sorts its marching orders, drawing sheet 11 from US 2026/0368089 A1
Sheet 11 How each column receives and sorts its marching orders
Checking memory addresses before letting tasks run, drawing sheet 12 from US 2026/0368089 A1
Sheet 12 Checking memory addresses before letting tasks run
Steps the chip follows to set up repeated loops, drawing sheet 13 from US 2026/0368089 A1
Sheet 13 Steps the chip follows to set up repeated loops
Mapping a three-dimensional block of data into memory, drawing sheet 14 from US 2026/0368089 A1
Sheet 14 Mapping a three-dimensional block of data into memory
Unpacking and sending work orders up each column, drawing sheet 15 from US 2026/0368089 A1
Sheet 15 Unpacking and sending work orders up each column
Staggering data delivery so every section arrives on time, drawing sheet 16 from US 2026/0368089 A1
Sheet 16 Staggering data delivery so every section arrives on time
Dividing on-chip memory between tasks to run them at once, drawing sheet 17 from US 2026/0368089 A1
Sheet 17 Dividing on-chip memory between tasks to run them at once
Patent filing page

Watchlists this filing belongs to

All watchlists →
Reader comments

No comments yet

Start the discussion

Real name or a handle, either is fine. Comments are read by a person before they appear, so allow a little time. Keep it about the filing.