Qualcomm Patents a Processor That Runs Two AI Calculations Simultaneously in One Step

By Patentlyze Team · Updated Jul 24, 2026

AI models do a lot of math, and right now that math often requires separate trips to different parts of a chip. Qualcomm is filing a patent to collapse two common operations into a single pass through one computing core.

How Qualcomm's unified math core speeds up AI

Imagine a factory assembly line where two separate stations do nearly identical jobs. Instead of routing parts through both, you redesign the line so one station handles both tasks. That's roughly what Qualcomm is doing here with AI chip math.

When your phone runs an AI feature, like generating a response or recognizing a photo, the chip performs enormous amounts of number-crunching involving grids of numbers called matrices. Two specific operations, multiplying entire grids together and multiplying matching individual cells, currently often require separate processing steps. Qualcomm's design folds both into a single core in one continuous operation.

The result is that the chip moves data around less, which is one of the biggest costs in computing. Less data movement typically means faster results and lower power use, both of which matter a lot on a phone or a wearable running on a battery.

How the GEMM kernel absorbs element-wise multiplication

The patent describes a computing core that handles two kinds of matrix math simultaneously rather than sequentially.

The first operation is element-wise multiplication (sometimes called a Hadamard product), where two grids of numbers are combined by multiplying each pair of cells that sit in the same position. The second is general matrix multiplication (GEMM), the backbone operation of nearly every neural network, where rows of one matrix are multiplied against columns of another in a specific pattern. GEMM is so fundamental that chip designers build dedicated hardware blocks just to run it fast.

Qualcomm's approach converts the results of element-wise multiplication into two intermediate representations stored in local (on-chip) memory, keeping the data close to the processor rather than sending it to slower off-chip RAM. Those representations are then fed directly into the GEMM kernel (the software routine that drives the GEMM hardware) to produce a final combined output.

• Process input matrices to produce two intermediate representations of element-wise products
• Store both representations in on-chip local memory
• Run the GEMM kernel on those representations in one pass to produce the final result

By keeping everything inside the same computing core and avoiding a separate hardware stage, the design cuts down on the data movement that normally adds latency and energy cost.

What this means for AI performance on Qualcomm chips

Nearly every modern AI model, from large language models to image classifiers, relies heavily on GEMM operations. Element-wise multiplication also appears constantly in transformer-style neural networks (the architecture behind most current AI). If Qualcomm can run both in a single core pass, that could translate to measurably faster inference (the step where an AI actually produces a result) and lower power draw on Snapdragon-powered phones and laptops.

For you as a user, this kind of low-level optimization is what eventually shows up as AI features that feel instant rather than sluggish, or as a phone that doesn't get hot while running on-device AI. It's not a splashy consumer announcement, but it's the kind of foundational chip work that determines whether on-device AI is practical at all.

Editorial commentary on a publicly published patent application. Not legal advice.