AMD Patents a Chip That Runs Two AI Systems in Sync to Score Similarities

By Patentlyze Team · Updated Jul 10, 2026

AMD has patented a chip design that runs two neural networks in parallel, using identical weights at exactly the same moment, so the chip can directly compare two images, sounds, or data samples without the usual back-and-forth.

What AMD's side-by-side neural network chip actually does

Imagine you're trying to find out if two passport photos are the same person. Today, AI systems typically run one photo through a neural network, then the other, and compare the results afterward. That two-step process adds time and complexity.

AMD's patent describes a chip that processes both photos at the same time, using two mirrored neural networks that share the exact same settings (called weights). Because both networks are in sync, the chip can make a direct, real-time comparison and spit out a score that says how similar the two inputs are.

The design is aimed at tasks where similarity comparison is the whole point: facial recognition, fingerprint matching, finding duplicate images, or even spotting fraud by comparing transaction patterns. Running it in dedicated hardware, rather than in software on a general-purpose chip, makes the whole process considerably faster.

How the systolic array processes both inputs at the same time

The patent centers on a siamese neural network (a type of AI architecture designed to compare two inputs by running them through identical, weight-sharing networks) built directly into silicon.

The key hardware component is a systolic array (a grid of small processors that pass data through in a wave-like pattern, commonly used in AI accelerators like Google's TPU). AMD's design packs two convolutional neural network (CNN) circuits into this array, one for each input. The critical detail: both circuits apply the same weight to their respective inputs simultaneously, not sequentially. Weights are the learned numerical parameters that define what a neural network has figured out during training.

After both inputs have been processed in parallel, a classifier circuit takes the two resulting representations and produces a similarity score. That score tells you how alike the two original inputs are.

The design avoids the common workaround of running one input, storing the result, then running the second input. Instead:

• Both CNN circuits share weight memory, cutting hardware duplication
• Parallel execution means the comparison is done in roughly half the time
• The classifier is co-located on the same device, avoiding data round-trips

What this means for face recognition and AI search hardware

Siamese networks are the backbone of a wide range of real-world AI applications: face verification, signature authentication, medical image comparison, and content-based image search. Right now, most of these systems run on general-purpose GPUs or CPUs, which were not designed with the two-input comparison pattern in mind. Dedicated hardware that handles both inputs at once could make these systems faster and more power-efficient.

For AMD, this fits into a broader push to compete in the AI accelerator market, where Nvidia currently dominates. A chip optimized for similarity-matching tasks could find a home in data centers running identity verification, security cameras, or large-scale media indexing, places where your face, fingerprint, or file gets compared against millions of stored records in milliseconds.

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