Qualcomm Patents a Lower-Power Way to Run a Core AI Calculation on Chips

By Patentlyze Team · Updated Jul 21, 2026

One of the most common calculations inside every AI model is also one of the most expensive to run on a chip. Qualcomm thinks it has found a way to cut the cost dramatically without wrecking the results.

What Qualcomm's cut-down AI math actually does

Imagine you're trying to split a pizza fairly among friends, and you need to figure out what percentage of the whole each slice represents. AI models do a version of this constantly, called softmax, to decide which of many possible answers is most likely. It's one of the most frequent calculations in any AI brain, and it's genuinely costly in terms of chip power and speed.

Qualcomm's patent describes a way to do that same math using rougher, simpler arithmetic, specifically by swapping out precise number formats for lower-bit representations and replacing expensive operations like exponents with good-enough approximations. Think of it like using a contractor's tape measure instead of a laser level: slightly less precise, but fast and totally adequate for the job.

The goal is to make AI run faster and use less power, especially on devices like phones or chips where you can't plug in a supercomputer. If the math is almost right and the chip runs cooler and longer, that's a real trade-off worth making.

How the four-step approximation pipeline works

The patent describes a four-step pipeline that replaces the standard softmax calculation with a cheaper chain of approximations.

• Step 1 (Exponent approximation): Instead of computing exact exponential values (a mathematically expensive operation), the system uses a non-uniform piecewise linear approximation, meaning it breaks the curve of the exponent function into small straight-line segments, each tuned to the part of the range where it applies. Non-uniform means the segments aren't all the same width; more segments are placed where the curve bends most, so accuracy holds where it matters.
• Step 2 (Normalization): A normalization pass adjusts the intermediate values so they stay in a workable numerical range, preventing the numbers from exploding or shrinking to zero.
• Step 3 (Inverse operation): The system computes a reciprocal (one divided by a number), which is the division step at the heart of softmax.
• Step 4 (Bitwise shift): Rather than a full multiply or divide, the final scaling is done with a bitwise shift, essentially moving binary digits left or right, which is far cheaper than a real multiplication on most hardware.

All of this is done using reduced bitwidth, meaning the numbers themselves are stored and processed with fewer binary digits than standard 32-bit floats, saving memory bandwidth and energy.

What this means for AI running on phones and chips

Softmax appears in nearly every transformer-based AI model, the architecture behind large language models and most modern image and speech AI. Running it efficiently is not a niche problem; it's one of the most-executed code paths in on-device AI. If Qualcomm can shave real cost from this single operation, the benefit multiplies across every token an AI generates.

This patent fits squarely into Qualcomm's strategy of making AI run well on Snapdragon chips inside phones and edge devices, where battery life and heat matter far more than in a data center. You'd feel this as longer battery life when using AI features, or as the ability to run a more capable model on the same hardware. It's the kind of invisible optimization that ends up mattering a lot in shipping products.

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