On November 16, 2026, NVIDIA announced plans with IBM and the U.S. Department of Energy's Fermi National Accelerator Laboratory (Fermilab) to build a co-design platform for quantum computers. At the same time, it released TCBP-JOBS, a decoder that estimates errors arising in quantum computation, in CUDA-Q QEC 0.9.0. The "up to roughly 13,000x" figure highlighted in the announcement compares the processing time for decoding simulation-generated data against conventional CPU-based software. The conditions differ between the time to decode a single instance and the time to process a large number of trials in bulk, and the two figures also mean different things for which stage of quantum computation can be accelerated.

AD

Using error-estimation history and searching in parallel on GPUs

In quantum error correction, TCBP-JOBS has the job of estimating the errors most likely to have occurred from measurement data. Quantum error correction protects a single piece of quantum information by encoding it across multiple physical qubits. The information obtained from the measurements used to detect errors is called the "syndrome," and the classical algorithm that uses it to estimate the errors to be corrected is called a "decoder."

In a research paper by Ben Howe and colleagues published by NVIDIA Research, the authors propose a new method that uses error-probability estimates to guide the search. The conventional combination used BP (Belief Propagation), which iteratively estimates error probabilities, together with OSD (Ordered Statistics Decoding), which searches according to a ranking of candidates. In this approach, the results of the iterative computation, such as the final or minimum BP values, are condensed into a single value. Error candidates are ranked on that basis, and correction candidates are searched following a single ranking.

TCBP, by contrast, feeds the history of how the estimates changed during the iterations into a small neural network. It derives adjusted error probabilities from this, and JOBS searches multiple candidates in parallel on the GPU while varying the ranking slightly. The system finally selects the most promising correction candidate. Its distinguishing feature is that it uses information from partway through the computation, which was previously condensed away, in the search.

The standard model has 529 learnable parameters. In the paper's standard evaluation setup, BP is iterated 10 times and JOBS searches 16 rankings. The model is trained separately for each type of code using simulation data, and the authors say training took about 10 seconds. Performance was also evaluated by processing simulation-generated inputs on CPUs and GPUs.

TCBP-JOBS searches for correction candidates within a finite range based on what it learned, so speed and accuracy vary with the number of candidates searched and the available memory. Performance also depends on the quantum error-correcting code and the training conditions.

Single decodes and bulk processing are compared under different conditions

Appendix B, Table 11 of the research paper, the basis for the "up to roughly 13,000x" figure, decodes the same 2,000 simulation trials one at a time with both methods and compares the average processing time. A trial here means one simulation of a circuit. The evaluation targeted Z-basis circuits of bivariate-bicycle codes, with the physical error rate of the SI1000 noise model set to 0.2%.

The comparison target, Tesseract, was run with a long-beam search setting and sparsification disabled, using one core of a Grace CPU. TCBP-JOBS, meanwhile, ran on one GPU in the same GB200 node. The CPU's average processing time used a value corrected to the speed at quiet times, when no other processes were running. However, not all trials were re-measured in a quiet environment; the correction is based on a ratio obtained by re-running some of the same trials.

The computing resources compared also differ between single decodes and bulk processing. The single-decode comparison pits one Grace CPU core against one GB200 GPU, whereas the bulk comparison pits a processing time assuming perfect distribution across 64 CPU cores against GPU batch processing time. Appendix C of the paper covers bulk processing for research use: running many trials to estimate the "logical error rate," the proportion of cases in which errors remain in the quantum information after decoding.

The main evaluation results are as follows.

Evaluation target CPU side GPU side Speedup in paper
Single decode, [[108,8,10]] One Grace CPU core, average 31.5 s One GB200 GPU, average 2.36 ms 13,300x
Single decode, [[144,12,12]] One Grace CPU core, average 53.5 s One GB200 GPU, average 4.09 ms 13,100x
Bulk processing, 28 circuits, 24.2 million trials total 12.1 weeks assuming perfect distribution across 64 cores One GB200 GPU, 16,384 trials at a time, about 45 minutes 2,720x

All use the SI1000 noise model with the physical error rate set to 0.2%. The single-decode evaluation uses the same 2,000 trials for each code, with CPU times corrected to quiet-time speeds. The bulk evaluation is a separate assessment that matches the per-circuit trial counts of a public benchmark, and it compares combinations of hardware and decoding method. The figures are based on Tables 11 and 13 of the paper.

The bulk evaluation used 28 circuits, including X-basis and Z-basis circuits of surface codes, color codes, and bivariate-bicycle codes. The CPU side's total computation was 130,257 thread-hours, the sum of each thread's run time. Assuming this could be distributed perfectly across 64 cores, the processing time was converted to 12.1 weeks. The GPU side's run time was 44.8 minutes, which Table 13 rounds to 45 minutes. Single-decode time bears on whether a result arrives in time for a quantum computer's next operation, whereas the bulk speedup contributes to research efficiency in comparing candidate error-correcting codes and quantum circuits.

On decoding accuracy, the two methods were broadly comparable under the batch evaluation conditions here. However, the relationship between speed and accuracy changes with the physical error rate. In Section 7 of the paper, the authors examined the physical-error-rate range of 0.01–0.05% for a 12-round Z-memory circuit of [[144,12,12]] using the SI1000 noise model, with a rare-event estimation method, and found that Tesseract achieved higher accuracy. In this low-error-rate evaluation, the splitting-method computation for estimating rare events was run independently eight times, and the median and spread of the results were obtained. This spread represents run-to-run variation and is not a confidence interval. When choosing a decoder, accuracy for the target code and physical error rate must be considered alongside processing speed.

AD

The co-design platform with IBM and Fermilab, and extensions to Slurm

The co-design platform NVIDIA plans with IBM and Fermilab will house NVIDIA DGX B200 systems alongside IBM quantum computers, including IBM Quantum Nighthawk. The plan is to design both quantum and classical computing to meet each other's requirements, researching everything from error-correcting codes and decoders to configurations that connect multiple quantum processors. The research targets also include applications in high-energy physics and nuclear physics. The fast decoding of simulation data by TCBP-JOBS can also be used when comparing multiple candidates in this co-design work.

Krysta Svore, NVIDIA's vice president in charge of quantum computing applications research, said of the purpose of the collaboration: "To make quantum computers useful for science, we need progress in error correction and applications, and in the systems that connect quantum processors with fast computing."

Supporting the operation of such computing environments is an extension to Slurm, which manages access to computing resources. Previously, each time a quantum processor (QPU) was introduced, developers had to build connection handling separately for each model and facility. The extended Slurm supports the QRMI and QDMI interfaces, allowing connection functionality developed by QPU makers to be reused at different sites. This makes it possible to manage access to QPUs in a unified way alongside CPUs and GPUs. Slurm 26.11 is scheduled to be available in late November 2026.

No specific start date for the co-design platform has been announced. As for the paper published by NVIDIA Research, we could not confirm its peer-review status, and we found no journal name or DOI, nor any independent replication. The authors list total response time, including the quantum control system, and suitability for real quantum computers as items for future evaluation.

To be used for real-time error correction on a quantum computer, a decoder must keep pace with the rate at which measurement data arrives and return results before the next quantum operation. Beyond average processing time, the distribution of latency, including cases where responses are slow, and the volume of data that can be processed continuously also matter. These performance characteristics will be the criteria for judging whether the speedup shown in bulk processing of research data can be applied to quantum error correction on real hardware.