Get the newsletter
AMD · AI / ML · Filed May 5, 2025 · USPTO published Aug 20, 2026 · Aug 20 release

AMD Patents a System That Splits AI Work Across Different Chips on the Fly

Most computers have a CPU, a GPU, and other chips sitting side by side, but AI workloads usually get routed to just one of them. AMD is patenting a way to break a single AI task into pieces and run each piece on whichever chip is best suited for it, right now, in the moment.

Block diagram of the heterogeneous computing system routing tasks between processors and combining results. Drawing from patent filing US 2026/0244503 A1.
Block diagram of the heterogeneous computing system routing tasks between processors and combining results.All 4 drawings ↓
Publication number
US 2026/0244503 A1(US20260244503A1)
Applicant
Advanced Micro Devices, Inc.
Filed · Published
May 5, 2025 · Aug 20, 2026
Inventors
Jipeng LI
Status when we published
Waiting for an examiner
Document
20 claims
US classification
718/102

How AMD's chip-splitting scheduler handles AI tasks

Imagine your laptop is trying to answer a question using an AI model. The GPU is already busy, but the CPU has spare capacity. Rather than making the AI wait for the GPU to free up, what if the computer could automatically split the job and send part of it to the CPU and part to the GPU, then stitch the answers back together?

That's what AMD is describing here. The system watches what computing resources are actually available at the moment a task arrives, then decides on the fly which calculations go to which chip. It's less about raw speed and more about not wasting what's already in front of you.

For you, the practical upside is that AI features running on your device could feel more responsive because the computer stops waiting around for one overloaded chip when another is sitting idle.

From the filing · CLAIM 1
… scheduling a first set of calculations to be made at a first type of processor of the heterogeneous computing system based on compute resources for the execution of the inference task on the first type of processor that are determined at runtime …

Translation: The system assigns tasks to specific chips based on what resources are available right then.

How the scheduler reads chip availability at runtime

The patent describes a dynamic scheduler sitting inside a heterogeneous computing system (a machine that has more than one type of processor, such as a CPU alongside a GPU or an AI accelerator). The scheduler's job is to intercept an inference task (the moment an AI model is asked to produce an answer or result) and decide, in real time, how to split the work.

The core mechanism works in two steps:

  • At runtime (while the computer is actually running, not beforehand), the scheduler checks how much compute capacity each processor type has available right now.
  • It then partitions the AI model's calculations: a first set goes to one processor type, a second set goes to another. When both sets are done, their outputs are combined into a single final result.

The patent emphasizes that the scheduling decision is made at runtime, not baked in during software design. That distinction matters because chip availability shifts constantly as other programs compete for resources.

The claim covers any heterogeneous computing system, meaning the approach could apply to desktops, laptops, or data-center servers that mix different processor architectures.

From the filing · THE ABSTRACT
At runtime, there may be various circumstances the computing system is faced with that makes the usage of one type of processor more optimum than another processor for a certain number or type of calculations.

Translation: Changing conditions decide which processor works best for any given task.

What this means for AI workloads on AMD hardware

AMD makes both CPUs and GPUs, and its newer chips tightly integrate multiple processor types on the same package. A runtime scheduler that intelligently routes AI inference work across those processors could reduce the bottlenecks that happen when a GPU is saturated but a CPU or AI accelerator is underused. For anyone running AI tools locally on an AMD-powered PC, the payoff would be AI responses that arrive faster without needing faster hardware.

The filing sits in a crowded area: every major chipmaker is working on ways to coordinate workloads across heterogeneous compute. Keeping tabs on where companies are placing their bets in chips and AI scheduling is exactly what the latest Big Tech patents covered on Patentlyze trace, and AMD's approach here of deferring the routing decision to runtime rather than compile time is the detail worth watching.

AMD's 31st filing we've tracked since May in our AI chip wars coverage, building on faster on-chip data formatting and a two-table math shortcut, shows the company continuing to push on how chips handle AI work.

Source: US 2026/0244503 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 AMD 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 · 4 sheets

4 drawing sheets from US 2026/0244503 A1 · click any drawing to enlarge

System architecture block diagram, drawing sheet 1 from US 2026/0244503 A1
Sheet 1 System architecture block diagram
Scheduler internal components, drawing sheet 2 from US 2026/0244503 A1
Sheet 2 Scheduler internal components
Inference execution flowchart, drawing sheet 3 from US 2026/0244503 A1
Sheet 3 Inference execution flowchart
Historical ratio tracking flowchart, drawing sheet 4 from US 2026/0244503 A1
Sheet 4 Historical ratio tracking flowchart
Patent filing page

More in AI training & infrastructure

All 324 AI training & infrastructure patents →

Watchlists this filing belongs to

All watchlists →