← Latest papers
⚛️ quantum physics

Scaling Pêlagos: Multi-GPU Orchestration for Relay Belief-Propagation Decoding

This paper presents an optimized multi-GPU orchestration of the Pêlagos relay belief-propagation decoder that achieves near-linear scaling and significant end-to-end speedups for large-scale quantum error-correction simulations on AMD Instinct supercomputers.

Original author: Xulong Tang, Dionysios Diamantopoulos, Tristan Müller, Apo Kayi, Nelson Mimura Gonzalez, Muhammad Osama, Seetharami Seelam, Haralampos Pozidis, Blake R. Johnson, Yasuko Eckert

Published 2026-11-13
📖 4 min read🧠 Deep dive

Original author: Xulong Tang, Dionysios Diamantopoulos, Tristan Müller, Apo Kayi, Nelson Mimura Gonzalez, Muhammad Osama, Seetharami Seelam, Haralampos Pozidis, Blake R. Johnson, Yasuko Eckert

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the author. For technical accuracy, refer to the original paper. Read full disclaimer

The dream of a fault-tolerant quantum computer relies on a single, stubborn hurdle: keeping information safe. Quantum bits, or qubits, are incredibly fragile, prone to flipping or vanishing due to the slightest environmental noise. To build a machine that can run complex calculations, scientists must create "logical qubits" that are protected by layers of error correction. This protection works by constantly checking for mistakes and fixing them before they spread, a process that requires a decoder to act as a rapid-fire referee. This referee must analyze a stream of error signals, known as syndromes, and decide almost instantly how to correct the system. The problem is that finding the right correction is a massive computational task. To design a decoder that is both fast enough for real-time use and accurate enough to handle the noise of a real machine, researchers must run many simulated scenarios across numerous operating points. Without a way to speed up these simulations, the design process stalls, leaving the path to useful quantum computing blocked by a wall of waiting.

A team of researchers from Advanced Micro Devices and IBM Research has found a way to break through that wall by rethinking how these simulations are run on modern supercomputers. They focused on a specific decoding method called relay belief-propagation, which is a heuristic algorithm designed to find corrections quickly without needing to calculate every single possibility. While this method is fast, running it on a single computer chip is not enough to explore the vast design space needed for future quantum computers. The researchers turned to graphics processing units, or GPUs, which are powerful chips capable of handling thousands of calculations at once. However, simply adding more GPUs to a single task often leads to diminishing returns. In their previous work, the team had built a system called Pˆelagos to run these decoders on GPUs, but when they tried to scale it up by connecting multiple chips to one computer, the system slowed down. The bottleneck was not the decoding itself, but the overhead of managing the work: the computer's main processor had to prepare data for each chip individually and then collect the results, creating a traffic jam that wasted the power of the extra chips.

To solve this, the team redesigned the orchestration of the work, moving away from a single computer program trying to manage multiple chips simultaneously. Instead, they assigned a dedicated, independent worker process to each GPU. Imagine a busy kitchen where a single head chef tries to direct eight cooks, handing them ingredients one by one and waiting for their plates before moving to the next; the head chef becomes the bottleneck. The researchers changed this to a model where each cook has their own station, prepares their own ingredients in parallel, and places their finished dish directly onto a shared counter for the final assembly. In technical terms, they separated the computer's main processor from the GPU management, allowing each chip to prepare its own data and write results directly to shared memory. This design eliminated the delays caused by the main processor trying to manage the chips, allowing the system to spend almost all its time actually decoding. The researchers also found that newer generations of chips, which have faster memory bandwidth, performed even better, handling the data movement required for these complex calculations more efficiently.

This work is not just a theoretical exercise; it represents a practical path forward for the development of quantum computers. The researchers deployed their system on a supercomputer at Oak Ridge National Laboratory as a pathfinding application for high-throughput QEC simulation. By making the simulation of error correction significantly faster, they allow scientists to test more decoder designs, explore more parameters, and find better solutions in a fraction of the time it would have taken before. This acceleration is critical for the "Genesis Mission," a broader effort to use high-performance computing and artificial intelligence to accelerate scientific discovery. The study confirms that by carefully managing how tasks are distributed across multiple processors, rather than just throwing more hardware at a problem, researchers can unlock the full potential of modern supercomputers. The findings suggest that the bottleneck in scaling quantum simulations is not the lack of raw computing power, but the software architecture used to direct it. With this new orchestration method, the path to evaluating the decoders needed for a fault-tolerant quantum future has been cleared, turning a years-long wait into a matter of weeks.

Drowning in papers in your field?

Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.

Try Digest →