Get the newsletter
Qualcomm · AI / ML · Filed Feb 13, 2025 · USPTO published Aug 13, 2026 · Aug 13 release

Qualcomm Patents a Method to Speed Up Image Data Processing for AI Training

Every time an AI model trains on images, those images have to be packed and unpacked at enormous speed. Qualcomm's new patent describes a compression method designed to do that unpacking in parallel, keeping expensive processors fed instead of waiting.

Computing device architecture incorporating processing units, memory controller, and system memory. Drawing from patent filing US 2026/0238834 A1.
Computing device architecture incorporating processing units, memory controller, and system memory.All 15 drawings ↓
Publication number
US 2026/0238834 A1(US20260238834A1)
Applicant
QUALCOMM Incorporated
Filed · Published
Feb 13, 2025 · Aug 13, 2026
Inventors
Amir Said
Status when we published
Waiting for an examiner
Document
26 claims
US classification
382/157

What Qualcomm's parallel image compression actually does

You're training an AI model on millions of images, and the bottleneck isn't the chips doing the math, it's the pipeline feeding them. Images have to be compressed so they don't eat all the memory, then decompressed just before the AI looks at them, and doing that one chunk at a time is slow.

Qualcomm's patent tackles this by processing many image chunks simultaneously instead of one after another. It uses a data structure called a range tree to decode compressed image data in large, parallel batches, which matches how modern chips (including Qualcomm's own) are designed to work.

The practical effect: AI training pipelines spend less time waiting on data delivery and more time actually learning. For companies running large training jobs, that kind of efficiency improvement adds up quickly across billions of image passes.

From the filing · CLAIM 1
… entropy decode, using parallel vector operations, the encoded vectors using vector-range tree coding (V-RTC) to generate decoded data; and reconstruct the decoded data into image blocks.

Translation: It decodes compressed image data simultaneously using a special method to prepare it for use.

How vector-range tree coding handles image blocks in parallel

The patent describes a decompression system built around a technique called Vector Range-Tree Coding (V-RTC). A range tree is a data structure (think of it as a lookup table organized like a family tree) that lets a processor quickly figure out what symbol or value an encoded bit sequence represents. V-RTC applies that tree lookup to whole vectors of data at once rather than one value at a time.

The pipeline the patent describes works in three main steps:

  • Retrieve: Pull multiple encoded image blocks from memory simultaneously.
  • Entropy decode: Run parallel vector operations through the range-tree structure to turn compressed bit sequences back into usable numbers. Entropy coding is a standard compression technique that assigns shorter codes to more common values.
  • Reconstruct: Reassemble the decoded numbers into full image blocks ready for the neural network.

One specific detail the patent highlights is converting signed integers (numbers that can be positive or negative) into unsigned integers (positive only) before decoding. This simplification lets the range-tree operations run more efficiently on parallel hardware, since unsigned arithmetic is cheaper for processors to handle in bulk.

The target application is high-throughput image processing, meaning scenarios where a system needs to decompress thousands or millions of images as fast as possible, which is exactly the pattern of large-scale neural network training.

From the filing · THE ABSTRACT
… convert signed integers to unsigned integers, decodes the unsigned integers using a range-tree structure, and reconstructs image blocks.

Translation: It converts numbers to a format that can be processed quickly in order to rebuild the images.

What this means for AI chips and training speed

For anyone using a device powered by Qualcomm chips to run on-device AI, the quality of that AI depends partly on how efficiently it was trained. Faster, cheaper training means companies can run more experiments, refine models further, or train on larger datasets without proportionally larger compute bills. The bottleneck this patent targets is real: data-loading and decompression are well-documented chokepoints in GPU and NPU training pipelines.

Qualcomm's angle here is particularly interesting because the company wants its chips to handle AI workloads not just in data centers but in phones and edge devices. A compression method that works efficiently on parallel hardware fits that roadmap. This filing sits alongside other chip-level AI efficiency patents among the interesting tech patents that show how semiconductor companies are optimizing the layers beneath the model itself.

Source: US 2026/0238834 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 Qualcomm 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 · 15 sheets

15 drawing sheets from US 2026/0238834 A1 · click any drawing to enlarge

Computing device system architecture, drawing sheet 1 from US 2026/0238834 A1
Sheet 1 Computing device system architecture
Three dimensional scene rendering pipeline, drawing sheet 2 from US 2026/0238834 A1
Sheet 2 Three dimensional scene rendering pipeline
Multimodal language model generation, drawing sheet 3 from US 2026/0238834 A1
Sheet 3 Multimodal language model generation
Hierarchical maximum value tree, drawing sheet 4 from US 2026/0238834 A1
Sheet 4 Hierarchical maximum value tree
Matrix grid maximum value examples, drawing sheet 5 from US 2026/0238834 A1
Sheet 5 Matrix grid maximum value examples
Two dimensional maximum value matrix, drawing sheet 6 from US 2026/0238834 A1
Sheet 6 Two dimensional maximum value matrix
Absolute coordinate matrix max values, drawing sheet 7 from US 2026/0238834 A1
Sheet 7 Absolute coordinate matrix max values
Coordinate grid index mapping tables, drawing sheet 8 from US 2026/0238834 A1
Sheet 8 Coordinate grid index mapping tables
Show the other 7 sheets
Qualcomm Patents a Method to Speed Up Image Data Processing for AI Training patent drawing, sheet 9
Sheet 9
Qualcomm Patents a Method to Speed Up Image Data Processing for AI Training patent drawing, sheet 10
Sheet 10
Qualcomm Patents a Method to Speed Up Image Data Processing for AI Training patent drawing, sheet 11
Sheet 11
Qualcomm Patents a Method to Speed Up Image Data Processing for AI Training patent drawing, sheet 12
Sheet 12
Qualcomm Patents a Method to Speed Up Image Data Processing for AI Training patent drawing, sheet 13
Sheet 13
Qualcomm Patents a Method to Speed Up Image Data Processing for AI Training patent drawing, sheet 14
Sheet 14
Qualcomm Patents a Method to Speed Up Image Data Processing for AI Training patent drawing, sheet 15
Sheet 15
Patent filing page

More in AI training & infrastructure

All 316 AI training & infrastructure patents →

Watchlists this filing belongs to

All watchlists →