Accelerating Quantum Error Correction Research with Pelagos and AMD Instinct GPUs
Nov 16, 2026
AMD powers the world's leading scientific computing, from the exascale supercomputers driving discovery in the United States and Europe to Lux, the first Genesis Mission system for AI-driven science. With leadership CPUs, GPUs, FPGAs, networking and open ROCm software, AMD gives nations, laboratories and enterprises a single open platform for high-performance computing, AI for science, sovereign AI, agentic AI and the path to quantum computing.
The fault-tolerant quantum computers of the near future will require advanced quantum error-correction (QEC) techniques. To develop and evaluate those QEC techniques, researchers increasingly depend on large-scale simulation workloads that can strain conventional computing resources. Faster simulation gives teams more opportunities to test ideas, refine decoder hyperparameters, and explore promising quantum codes and decoding algorithms.
To help researchers run these studies more efficiently, IBM and AMD developed Pêlagos, an open-source software tool for accelerating large-scale QEC decoder simulations. Now, IBM and AMD are releasing Pêlagos to enable and accelerate efforts to develop and evaluate new decoding strategies, to study existing decoders in greater depth, and to expand the toolbox available to QEC researchers. Anyone interested in trying Pêlagos for themselves can explore the code and a detailed tutorial on GitHub, or learn more about the project’s design, implementation, and scaling results on AMD Instinct™ GPUs via the new joint IBM and AMD white paper.
Decoder simulation presents a natural opportunity to take advantage of the parallelism offered by modern GPU accelerators. However, high-performance GPU kernels can be difficult for decoder algorithm researchers to write, maintain, and adapt, even in this age of agentic coding tools. To address this challenge, the Pêlagos team used the Python-based Triton framework to build a high-throughput implementation of Relay-BP, a promising QEC decoder designed for quantum low-density parity check (qLDPC) codes, that is also easy to read, manipulate, and adapt to explore new research directions in decoder algorithm design. The name Pêlagos comes from the Greek word for “open sea.”
Giving researchers more room to explore
Quantum error correction protects logical information by encoding it across physical qubits. A decoder interprets measurements called syndromes to infer which errors may have occurred. Evaluating a decoder requires repeating this workflow across many sampled error patterns drawn from a simulated noise model, then measuring how often the protected information is recovered correctly.
The challenge grows as researchers compare different codes, noise models, decoder parameters, or even decoder algorithms. As researchers push toward the exceptionally low logical error rates needed for fault-tolerant quantum computing, special care is needed in the statistical evaluation. Faster decoder simulations can increase the number of samples processed within a research budget, helping teams investigate more configurations and evaluate decoder performance with greater statistical certainty.
Quantum low-density parity-check (qLDPC) codes, and particularly bivariate bicycle codes, are an important target for this work. QEC researchers find them compelling because they have the potential to protect quantum information with fewer physical qubits than other approaches. However, their decoding requirements create substantial computational work. Pêlagos accelerates the simulation needed to study that trade-off.
Adaptive software for demanding decoding workloads
At the heart of Pêlagos is a GPU-accelerated implementation of Relay-BP, a quantum error-correction decoder adapted from belief propagation (BP), a family of algorithms used widely in classical LDPC decoding. Relay-BP was purpose-built for qLDPC codes, which present unique decoding challenges. It works by iteratively exchanging information about possible errors using successive decoding attempts, like legs in a relay race, to improve its chances of finding a valid solution.
To accelerate Relay-BP on GPUs, Pêlagos processes many independent syndromes together, allowing multiple decoding problems to be evaluated in parallel. However, it takes more than a high degree of parallelism to create a high-throughput implementation of Relay-BP. In tests reported in the Pêlagos paper, a direct port of Relay-BP delivered less throughput than a single CPU core, while Pêlagos's optimized kernels reached throughput rates up to 97× higher. Achieving that level of performance requires highly efficient decoding kernels that can adapt to the diverse compute and memory resources that are available in modern GPU architectures. See Pêlagos paper, Figure 13.
The Triton framework provides a developer-friendly solution to this problem by combining a NumPy-like interface for expressing tensor computations with a tuning system that automatically discovers implementation details like memory layouts and workload partitioning specific to each GPU. Pêlagos further extends this capability with device fingerprinting and cached tuning results. This enables Pêlagos to deliver strong performance across GPU vendors and models while still being easy to read, modify, and adapt for future research.
IBM has deployed Pêlagos at cluster scale to dramatically speed up decoder research. Earlier this year, a 312-experiment campaign launched on IBM’s BlueVela cluster using 1,024 GPUs finished in under 20 days. Estimates show the same campaign would take roughly 45 years to execute on a single GPU. This campaign illustrates the scale of the simulation tasks that power modern decoder research, and it provides the foundation for the Pêlagos team’s work on AMD multi-GPU systems. See Pêlagos paper, Figure 18.
Helping multiple AMD GPUs work efficiently together
The new joint IBM and AMD white paper explores how Pêlagos scales on multi-GPU compute nodes with AMD Instinct™ GPUs. In this context, the challenge is that although independent decoding tasks provide ample parallel work, preparing inputs and collecting outputs can become a bottleneck as more GPUs are added. Adding accelerators delivers more value when the surrounding software keeps pace.
AMD and IBM solved this problem through careful partitioning and distribution of the workload across CPUs and GPUs. In particular, the team’s design isolates problem preparation on the CPU so that persistent worker processes on each GPU have a consistent stream of tasks, while host shared memory collects results without stalling the GPUs.
For the one-gross (144-qubit) bivariate bicycle code benchmark, moving from one to eight GPUs increases optimized end-to-end throughput by 5.68× on AMD Instinct MI300X GPUs and by 6.23× on the MI355X GPUs. At the same eight-GPU count, the optimized workflow delivers 2.64× the original Pêlagos workflow’s throughput on the MI300X system and 2.07× on the MI355X system. These comparisons show both the benefit of additional GPUs and the benefit of reducing orchestration overhead. These are workflow comparisons within each platform, not a performance comparison between the two GPU models. See companion white paper, Figure 2 and Table III.
For quantum researchers, that means more decoding work completed within a given simulation study, and more opportunities to explore decoder performance across a broader set of configurations. Note that these results measure aggregate simulation throughput. Live quantum hardware has separate requirements for per-syndrome response time.
Connecting quantum research with AI and HPC infrastructure
The MI355X evaluation runs on the AMD Lux supercomputer at Oak Ridge National Laboratory, showing how an AI and high-performance computing system can support quantum error-correction research. This deployment provides a concrete quantum-computing use case for the infrastructure supporting the Genesis Mission's scientific-discovery goals. The evaluated platforms combine AMD Instinct GPUs with AMD EPYC™ processors and run PyTorch on AMD ROCm™ software.
For AMD, this work serves as a concrete example of how classical infrastructure supports quantum progress. Accelerated decoding helps researchers evaluate candidate designs before deployment, while portable software gives them a way to run those studies across different computing platforms. Collaboration between decoder developers and system specialists helps translate those capabilities into useful application performance.
Pêlagos is jointly maintained by IBM and AMD engineers and includes a self-tuning GPU decoder with multi-GPU support, integration with the Stim simulator, and step-by-step tutorials for getting started. Researchers interested in trying Pêlagos can explore the code and tutorials on GitHub, experiment with their own error-correcting codes and decoders, and contribute back to the project.
To learn more, explore the original Pêlagos paper for details on the adaptive decoder architecture, and read the more recent companion white paper for the multi-GPU implementation and performance results on AMD Instinct™ systems. Together, they offer practical guidance for teams building the next generation of quantum error-correction simulation workflows.
Explore the research
- Pêlagos code and tutorial on GitHub
- Pêlagos paper — An Adaptive GPU Decoder for Error Correction Simulations of Quantum LDPC Codes
- Companion white paper — Scaling Pêlagos through Multi-GPU Orchestration
- AMD and the hybrid future of quantum computing
- IBM announcement — ORNL, IBM, AMD launch co-design platform for QEC