Pêlagos achieves near-linear GPU scaling.

Near‑Linear GPU Scaling for Quantum Error‑Correction

Researchers from Advanced Micro Devices (AMD) and IBM Research have shown that the Pêlagos relay‑belief‑propagation decoder can scale almost linearly when run on multiple AMD Instinct GPUs. By redesigning the orchestration layer to give each GPU its own process and HIP context, they eliminated host‑side bottlenecks that typically limit multi‑GPU performance.

Multi‑GPU Orchestration Improves Throughput

The new design isolates data‑shard preparation on the host and uses shared memory for returning correction arrays. This approach lets each GPU work on its own “shard” of syndrome data without costly inter‑process communication. On an eight‑GPU node, the team achieved:

• One‑gross [[144,12,12]] code: 5.68× speed‑up on MI300X and 6.23× on MI355X GPUs.
• Two‑gross [[288,12,18]] code: up to 7.58× (MI300X) and 7.45× (MI355X) speed‑up.

These figures are measured using GPU‑only decoding time, showing that adding GPUs increases the number of syndromes processed concurrently with only a modest rise in per‑GPU workload.

Practical Deployment on a Supercomputer

The decoder was deployed on Lux, Oak Ridge National Laboratory’s AMD Instinct MI355X‑based supercomputer, as a “path‑finding” application for high‑throughput quantum error‑correction (QEC) simulation. The implementation kept the original single‑GPU kernels unchanged; all scaling gains come from the surrounding node‑local preprocessing and output collection.

Implications for Large‑Scale QEC Studies

Large QEC codes generate massive detector graphs (e.g., the two‑gross circuit has over 1 million non‑zero matrix entries). Efficient multi‑GPU scaling reduces the time required to explore parameter spaces, tune memory‑strength intervals, and evaluate logical error rates. The authors report no logical failures at relay depth R = 50 after adjusting the memory‑strength interval, indicating that the decoder maintains high fidelity while scaling.

Open Source and Reproducibility

The source code for the multi‑GPU orchestration is publicly available (commit 0127224), enabling other groups to verify the results or build upon them. While the paper provides detailed hardware and software configurations, independent replication beyond the authors’ own tests has not yet been reported.

Bottom line: The work demonstrates that existing HPC infrastructure can be leveraged to accelerate QEC simulations, bringing the community a step closer to practical, large‑scale quantum error correction.

--- For more technical details, see the pre‑print by Xulong Tang et al. on arXiv.