IBM announced Pelagos as open-source software at SC26 for evaluating quantum error-correction decoders. The project implements GPU-accelerated Relay-BP decoding for qLDPC codes; a paper submitted on November 12, 2026, reports eight-GPU results that separate whole-workload speedups from gains in the decoding stage alone. (IBM’s announcement; Pelagos repository; scaling paper)

For a useful performance comparison, keep the decoding-stage speedup separate from the speedup for the full run.

What Pelagos does

Pelagos is a Triton-based GPU decoder for quantum low-density parity-check (qLDPC) codes. The project repository describes its implementation as Relay-BP-S: min-sum belief propagation with disordered memory strengths. That makes Pelagos a tool for simulating and evaluating a specific class of quantum error-correction decoders, not a quantum processor. (Pelagos project documentation)

The repository lists the installation command pip install pelagos-decoder[gpu]. It describes GPU targets for NVIDIA CUDA, AMD ROCm/HIP and Intel SPIR-V through Triton. Decoding requires a GPU; the project has no CPU fallback and directs CPU users to the separate Rust relay-bp implementation. Its documented multi-GPU execution is limited to one node. (Pelagos repository)

What the eight-GPU results measure

The scaling paper reports tests using eight AMD Instinct MI300X GPUs and eight AMD Instinct MI355X GPUs. It evaluates a one-gross circuit, [[144,12,12]], and a two-gross circuit, [[288,12,18]]. The reported speedups vary by circuit and by whether the timing covers the full run or GPU decoding alone. (Scaling paper)

Circuit and metricEight AMD Instinct MI300X GPUsEight AMD Instinct MI355X GPUs
One-gross [[144,12,12]], end-to-end5.68×6.23×
Two-gross [[288,12,18]], end-to-end6.92×7.01×
Two-gross [[288,12,18]], GPU decoding time7.58×7.45×

For the two-gross circuit, 7.58× on MI300X and 7.45× on MI355X describe GPU decoding time. The corresponding end-to-end results are 6.92× and 7.01×. Since end-to-end timing includes more than the decoding stage, the roughly 7.5× figures do not describe the speedup of the full run. These are results reported in the paper, not independent replications. (Scaling paper)

Pelagos on Lux

The paper also reports deploying Pelagos on Lux, an Oak Ridge National Laboratory supercomputer based on AMD Instinct MI355X. This is separate from Umbra, another AMD–IBM–ORNL effort discussed in NeoTeo’s earlier coverage of Scott Tease’s view of AI and HPC.