Research

NVIDIA, IBM and Fermilab Plan Modular Quantum Platform

NVIDIA, IBM and Fermilab proposed a modular quantum platform, while a GPU decoder posted a benchmark-specific 13,000-times speedup and Slurm plans native QPU scheduling.

NVIDIA, IBM and Fermilab Plan Modular Quantum Computing Platform
NVIDIA, IBM and Fermilab Plan Modular Quantum Computing PlatformDarkoStojanovic / pixabay
By Sophie Lindqvist5 min read

Why it matters

  • On November 16, 2026, NVIDIA reported up to a 13,000-times decoder speedup in one benchmark comparison.
  • TCBP-JOBS cut one research computation from 12 weeks to 45 minutes, according to NVIDIA.
  • IBM Quantum Nighthawk supports up to 7,500 two-qubit operations on 120 qubits at up to 100,000 samples per second.
  • TCBP-JOBS recorded 10 to 7,400-times serial speedups against Tesseract’s sparsified presets; logical error rates ran up to 4.5 times lower on the largest codes.
  • Slurm 26.11 targets later November 2026, with QRMI and QDMI support and six named adopters.

By Sophie Lindqvist

NVIDIA reported on November 16, 2026, that its newly released GPU-accelerated quantum error-correction decoder achieved up to a 13,000-times speedup in one benchmark comparison. NVIDIA Research’s abstract limits that figure to unsparsified long-beam Tesseract on bivariate-bicycle codes. Against Tesseract’s sparsified presets, serial speedups ranged from 10 to 7,400 times.

The result has two practical stakes. Fermilab chief technology officer and SQMS director Anna Grassellino said the initiative “would help establish the capabilities and research community needed for a future Department of Energy quantum computing user facility.” NVIDIA says TCBP-JOBS reduced one research computation from 12 weeks to 45 minutes, accelerating the evaluation of candidate codes and decoder configurations.

What exactly are NVIDIA, IBM and Fermilab proposing?

NVIDIA said it plans to develop a modular quantum-classical co-design platform with IBM and Fermi National Accelerator Laboratory, or Fermilab. Fermilab describes the initiative as a proposal, not an operating facility. The distinction matters because the announcement describes planned work rather than a deployed facility.

The initial design would physically co-locate IBM quantum computers with NVIDIA DGX B200 accelerated-computing systems. That setup would include computers powered by IBM Quantum Nighthawk.

The plan couples quantum devices with their classical control systems and scientific software. Researchers would test codes, decoders and applications against shared benchmarks to see how those components behave together.

Researchers would develop and evaluate error-correction codes, decoders and quantum-classical applications. They would also study how classical systems scale alongside multiple linked quantum processors. For now, this is a research proposal, not an operating facility.

Why does Fermilab matter to the research program?

IBM describes Nighthawk as “its most powerful quantum processor.” It supports circuits of up to 7,500 two-qubit operations on 120 qubits and speeds of up to 100,000 samples per second.

The scientific program would concentrate on high-energy and nuclear physics. Fermilab would lead engagement with researchers and coordinate software workflows and access through HEPCloud.

IBM, NVIDIA and Fermilab would develop algorithms, shared software and benchmarks for lattice gauge theories, neutrino interactions and nuclear simulations. Those problems would translate scientific needs into requirements for future fault-tolerant quantum systems. Fermilab supplies both the scientific use cases and a path toward shared Department of Energy access.

How would modular quantum systems connect?

Modularity is intended to connect processors for calculations too demanding for a single device. Researchers would first model and emulate communication between modules.

They would study link errors, latency and feedback, then use the results to guide IBM and Fermilab’s Superconducting Quantum Materials and Systems (SQMS) Center work on physical interconnects. That sequence tests communication behavior before the partners build more specialized hardware.

In a later phase, SQMS intends to integrate ultracoherent superconducting cavities. The team would investigate their use as quantum memories, entanglement buffers and resources for bosonic error correction. The network tests come first and will set the requirements for later hardware.

What does the 13,000-times decoder result actually cover?\n

TCBP-JOBS addresses the classical side of quantum error correction and is available separately through CUDA-Q QEC 0.9.0. Error correction encodes quantum information across multiple physical qubits; a decoder interprets measurement data to identify likely errors so a system can correct them.

NVIDIA says the algorithm reduced one research computation from 12 weeks to 45 minutes. That shortened the evaluation of candidate error-correction codes and decoder configurations.

The method combines trajectory-calibrated belief propagation with jittered OSD beam search. The first technique learns calibrated fault probabilities from belief-propagation trajectories. The second uses those probabilities to guide parallel GPU searches, making repeated code evaluation faster. NVIDIA Research’s abstract sets the benchmark boundaries:

  • The 13,000-times result compares TCBP-JOBS with unsparsified long-beam Tesseract on bivariate-bicycle codes.
  • The comparison uses logical-error-rate ratios of 0.97 to 1.57.
  • Serial speedups against Tesseract’s sparsified presets range from 10 to 7,400 times.
  • Tests cover surface, color and bivariate-bicycle code memory and surgery circuits.
  • Logical error rates were comparable to Tesseract and up to 4.5 times lower on the largest codes.
  • Batched execution supports offline simulation.
  • Serial memory and logical-surgery experiments show millisecond-scale latency for QPUs with longer logical clock cycles.

These results do not establish a uniform speedup across decoders or workloads. The 13,000-times label therefore covers one comparison, while the broader case rests on faster code evaluation, comparable error performance and scheduling features tested in defined settings. The headline number covers one benchmark, not every decoder or workload.

How will Slurm schedule shared quantum hardware?\n

Slurm previously had no native concept of a quantum processing unit, or QPU. Research centers attached QPUs through plugins outside the scheduler’s core resource model.

IBM, NVIDIA and Pasqal worked with the open-source community on the Quantum Resource Management Interface, or QRMI. That vendor-neutral layer connects to Slurm Hierarchical Resources.

According to NVIDIA’s Slurm announcement, GPUs operate within individual computing nodes, while a QPU can be shared across multiple nodes. Those different shapes make quantum allocation a site-level problem rather than a simple node-local reservation.

Hierarchical Resources lets administrators manage the shared QPU, allocate and reserve its time, and account for its use under familiar high-performance-computing policies. A hybrid workload can run as a normal Slurm job without a separate queue or login.

NVIDIA says support for the QRMI and QDMI quantum-HPC interfaces will let QPU builders reuse one integration across deployments. Builders would no longer need bespoke software for every QPU-center combination.

The listed adopters are Alice & Bob, IBM, IQM, Pasqal, Quantinuum and MQSC. Slurm 26.11 is targeted for availability later in November 2026. Slurm aims to manage quantum hardware like other shared computing resources, though the release remains a target.

What has to happen next?

The proposal and software releases address three separate constraints: evaluating error correction, scheduling shared hardware and connecting processors. The first interconnect studies form the central hardware test.

Their measurements of link errors, latency and feedback should set the requirements for later fault-tolerant, multi-processor systems before Fermilab adds its cavity hardware. Until those studies produce data, the larger platform remains a proposal and the headline decoder gain remains benchmark-specific. The next checkpoints are the interconnect models, the platform’s error-correction evaluations and the planned Slurm 26.11 release.

Source: unite.ai

Share this article:

More from Sophie Lindqvist

Sophie Lindqvist

Show full bio

Staff writer covering regulation, policy and their impact on artificial intelligence and machine learning at AI In Context.

749 articles

Related articles

  1. IBM's Nighthawk r3 hits 520 qubits and 19x speedup
  2. NVIDIA Open-Sources Run:ai Under Apache 2.0
  3. NVIDIA to invest up to $100 billion in OpenAI for 10 gigawatts of AI infrastructure
  4. NVIDIA Ships Nemotron 3.5 Super VL, a 120B Open Multimodal Agent Model
  5. NVIDIA's Nemotron for Science pairs open model with chemistry bench

« Previous articleNext article »