NVIDIA, IBM and Fermilab plan quantum codesign platform | Magica
NVIDIA, IBM and Fermilab plan modular quantum platform as decoder benchmarks show trade-offs
Editorial Team
••📖8 min read
The partners plan to combine IBM quantum processors, NVIDIA computing and Fermilab physics expertise to study modular quantum systems. A separate decoder release promises faster error-correction research, with important limits on its performance comparisons.
IBM quantum computers and NVIDIA DGX B200 systems would share a research platform focused on modular quantum computing and physics applications.
NVIDIA separately released an error-correction decoder claiming up to 13,000 times faster decoding in research workloads.
The technical comparison uses different computing resources; the proposed platform has no announced start date or cost.
Accelerated-computing supplier NVIDIA, quantum-computer maker IBM and the U.S. Department of Energy’s Fermi National Accelerator Laboratory announced plans on November 16, 2026, for a platform to design quantum computers and their supporting classical systems together. Researchers would test error correction, scientific applications and communication between quantum processors to guide future modular hardware.
The practical issue is how much conventional computing, networking and control quantum machines will need as they grow. NVIDIA supplies accelerated computing and quantum software; IBM contributes superconducting quantum processors and networking expertise; Fermilab supplies particle-physics problems and superconducting hardware research. Their work could help define the requirements for future scientific computing facilities, but the announcements describe a research program rather than a completed fault-tolerant computer.
Conceptual hardware illustration for the proposed platform: IBM Quantum System Two, a Fermilab nine-cell superconducting cavity and NVIDIA DGX B200. Credits: IBM; Ryan Postel, Fermilab; NVIDIA. Source: Fermilab.
Connecting processors starts with emulating the links
The proposed platform would physically co-locate NVIDIA DGX B200 accelerated-computing systems with IBM quantum computers, including machines powered by IBM’s Nighthawk quantum processor. NVIDIA describes hardware colocation as under way; IBM and Fermilab frame the modular platform itself as planned. IBM says Nighthawk can run circuits containing up to 7,500 two-qubit operations on 120 qubits, at speeds up to 100,000 samples per second, in its announcement. These are processor capabilities, not a demonstration that the proposed interconnected system works.
South Korea is strengthening device checks and considering public Wi-Fi shutdowns before its November 19 college entrance exam, after a tax-accountant exam top scorer admitted using AI to cheat. The response still depends heavily on human screening, with no electronic device detectors planned for this year's test.
Editorial Team
Modularity means making multiple quantum processors cooperate on calculations that exceed a single device’s capacity. Initially, researchers would model and emulate communication between modules, measuring the effects of link errors, delays and feedback. Those results would guide development of physical quantum interconnects, according to Fermilab’s plan.
Fermilab is the national laboratory for particle physics and accelerator research, managed for DOE by Fermi Forward Discovery Group. Its Superconducting Quantum Materials and Systems Center, or SQMS, applies the laboratory’s experience building particle accelerators to superconducting quantum hardware. It already collaborates with IBM on quantum technologies and interconnects. The center is one of DOE’s five National Quantum Information Science Research Centers and includes more than 40 partner institutions.
NVIDIA’s software also has a history at the laboratory. A report on the announcement recounts Fermilab’s September statement that NVIDIA’s CUDA-Q Logical software layer cut a design job from five months to three weeks. That is a separate, previously reported workflow improvement, not a result from the proposed platform.
In a later phase, SQMS intends to add its superconducting cavity hardware. Researchers would investigate those cavities as quantum memories, buffers for entanglement and resources for bosonic error correction. Fermilab would also coordinate researcher access and software workflows through its HEPCloud infrastructure and bring in high-energy and nuclear physicists to develop algorithms and benchmarks for particle interactions, neutrino physics and nuclear simulations.
The choice of laboratory therefore matters to the design process: scientific users would help set hardware requirements. This approach also has an alternative using AMD’s classical computing hardware. IBM’s November 5 announcement describes AMD computing as the foundation for error-correction research, decoding optimization and software orchestration in its platform with Oak Ridge National Laboratory’s Quantum Science Center. Fermilab says that recently launched platform also targets materials applications and would complement the proposed Fermilab collaboration, with potential exchanges of software and benchmarking methods.
What the decoder speedup measures
Separately, NVIDIA released TCBP-JOBS in its CUDA-Q QEC 0.9.0 library for quantum error correction. Errors limit how long and how complex quantum calculations can be. Error correction spreads quantum information across multiple physical qubits; a classical decoder interprets measurements to identify likely errors. NVIDIA claims decoding speedups of up to 13,000 times in research workloads in its release. It also gives an example of design-research computation falling from 12 weeks to 45 minutes; those durations describe a different ratio from the headline maximum.
The mechanism combines a small neural network that estimates error probabilities from the history of a belief-propagation algorithm with a GPU-parallel search for corrections. NVIDIA’s technical guide supplies essential limits on how to interpret the performance evidence.
Its displayed comparison covers 14 simulated memory circuits spanning surface, color and bivariate-bicycle error-correction codes, at a physical error rate of 0.002 in the Z measurement basis. It compares TCBP-JOBS on one NVIDIA GB200 with the long-beam configuration of Google Quantum AI’s Tesseract decoder, a CPU alternative requiring neither a GPU nor a fitted model. The Tesseract timings come from repository results rather than a fresh run; the recorded data omit the CPU hardware and decoder version. Faster Tesseract configurations that trade accuracy for speed are outside that displayed comparison.
NVIDIA explicitly describes the latency chart as one CPU thread versus an entire GPU, measuring time to an answer rather than performance on equivalent hardware. The GPU decoder stays below a millisecond per error-measurement round; Tesseract is faster on the smallest surface-code case but takes several seconds per round on larger bivariate-bicycle codes. Across the 14 circuits, accuracy is statistically indistinguishable at the reported sample size on 12; TCBP-JOBS has fewer failures on the distance-11 color code and more on the distance-9 color code. These are independently sampled results from the same distributions, not tests on identical shots.
For research throughput, NVIDIA reports decoding 12 million shots in 0.36 GB200 GPU-hours, compared with 63,334 CPU core-hours for the Tesseract repository results covering the same number of independently sampled shots. Those totals use different resources and timing modes. The GPU figure excludes sampling, detector-model construction and neural-network fitting. They show the potential to accelerate large simulation studies, but do not establish a comparable reduction in financial cost or real-time error-correction latency.
The default learned model is small: NVIDIA reports 529 trainable weights and fitting in under a minute in its GPU development runs. Still, the guide says changing code families or belief-propagation settings requires validating or refitting it. Its bounded search does not guarantee the most likely correction. These limits matter when moving from a simulated research workload to a different device or noise regime.
Company-reported compute totals for 12 million independently sampled shots per decoder, using different resources and timing modes. These are not hardware-normalized speed or cost comparisons. GPU time excludes sampling, detector-model construction and model fitting. Source: NVIDIA CUDA-Q QEC documentation.
A route through existing supercomputing infrastructure
The partners also aim to make quantum resources easier to use through Slurm, the open-source software that schedules supercomputer jobs. IBM says quantum processors previously had to be attached through plugins outside Slurm’s core resource model. They can be shared across multiple computing nodes, so scheduling must account for that shared access.
IBM says it worked with NVIDIA, Pasqal and the open-source community on the open-source, vendor-agnostic Quantum Resource Management Interface linked to Slurm’s Hierarchical Resources capabilities. This allows centers to allocate, reserve and account for shared quantum processors using familiar scheduling policies, without a separate queue or login. The announcement report notes that NVIDIA did not give a readiness date for the Slurm features.
NVIDIA has a stake in this layer beyond selling GPUs. It acquired SchedMD, Slurm’s leading developer, while committing to keep Slurm open source and vendor neutral. Its role now spans both accelerated hardware and development of software that allocates computing resources. IBM’s vendor-agnostic interface is a material qualification: the integration is intended to support quantum processors from different suppliers.
The announcements also point to existing hybrid scientific work. IBM reports a 79-fold accuracy improvement over earlier runs of the same iron–sulfur-cluster experiment with Japan’s RIKEN research institute and NVIDIA, approaching a leading classical method. The work used circuits on 72 qubits, collected 5 million samples and combined quantum computing with GPU calculations. It is a comparison with the team’s earlier experiment, not a claim of 79-fold superiority over classical computing.
NVIDIA separately highlights calculations of binding energies for two protein–ligand complexes containing up to 12,635 atoms, compared with a previous 303-atom hybrid benchmark. Those calculations divided the systems into fragments and combined quantum sampling with GPU computing; the atom count does not describe a wholly quantum simulation. NVIDIA says the work is a Gordon Bell Prize finalist. Japan’s New Energy and Industrial Technology Development Organization, or NEDO, supports both research efforts. These are distinct application studies, not results produced by the proposed Fermilab platform or proof of its modular architecture.
The next test is a working system with measurable costs
The partners have not disclosed a start date or cost for the platform, a gap also noted in the report on the announcement. Fermilab presents a future DOE quantum computing user facility as a long-term ambition, rather than an approved outcome of this collaboration.
The next meaningful evidence will be results showing how simulated links translate into physical interconnects, how error correction performs within an integrated system, and what classical resources scientific users actually require. For the decoder, comparisons that specify hardware, software versions, accuracy and full workflow costs would clarify how much of the research speedup carries into deployment.
Russia’s objections to observer participation stalled the opening of a Geneva weapons conference, where governments must decide whether to authorize negotiations on autonomous weapons and whether any resulting rules should be binding.
Apple’s iOS 27.3 beta ties higher allowances for server-based AI to existing iCloud+ plans. Basic Siri commands remain free, but undisclosed daily limits leave the value of paid access uncertain.
Federal Reserve Chairman Kevin Warsh outlined AI scenarios spanning shared prosperity, job displacement and an investment downturn. The exercise offers no forecast or rate guidance as colleagues worry that the infrastructure buildout is adding to inflation.
OpenAI is offering approved organizations a free classifier for suspected child sexual abuse images, including AI-generated material. Early results suggest less review work, but existing detection alternatives and the need for human review qualify the launch.
Nvidia plans scientific computing racks for early 2027, extending Vera beyond its initial AI deployments. Its 2.4× performance-per-watt claim comes from one airflow simulation against AMD Turin; performance against newer rivals and delivered costs remain unresolved.
OpenAI’s CEO expressed empathy for Nvidia founder Jensen Huang but withheld judgment on reinstatement. His confidence in continued AI financing comes with a direct commercial stake and unresolved deployment risks.
Google’s 8,192-chip Ironwood submission completed the MLPerf DeepSeek-V3 benchmark in 1.70 minutes, close to NVIDIA’s 1.53 minutes at the same accelerator count. Google’s scaling data also shows the extra chip-time required for its fastest result, leaving cost and energy comparisons unresolved.
TensorWave will supply dedicated AMD infrastructure under a six-year OpenAI agreement, with delivery expected from 2027. Its operating record and financial backing sit alongside an undisclosed contract price and a recently terminated site lease.
California alleges OpenAI endangered the public and misled consumers by testing powerful AI models without adequate controls. The lawsuit follows a Hugging Face intrusion that OpenAI’s own account traced to containment failures, missing monitors and mishandled warning signs.
Perplexity has launched Frontier Research to develop AI that can formulate questions, test hypotheses and produce interpretable discoveries. The company promises more than 10,000 GPUs, but existing automated-research systems show why computing resources and generated papers alone cannot establish scientific value.
The new lab led by Stanford professors Yejin Choi and Carlos Guestrin has backing from Nvidia, Cisco and Snowflake. It promises systems that improve on company work, but research benchmarks leave production costs and customer control unproven.
NVIDIA announced a Rubin HPC GPU with more than 50,000 CUDA cores, sparse data acceleration and seventh-generation NVLink. The design targets simulations used by scientists and AI agents, but missing performance, power and pricing details leave its practical advantage unproven.
OpenAI will open 50,000 new ChatGPT access grants on November 23, covering up to 150,000 researchers and collaborators. New recipients get Plus instead of Pro, with first-come selection replacing a lottery. Its pilot study reports perceived time savings but does not establish faster discovery.
OpenAI reports production healthcare workflows at CVS and Cigna and stronger computer-use evaluations. The results show progress across disconnected software, but leave complete-task reliability and total deployment costs unresolved.
President Donald Trump supported a return for Nvidia co-founder Jensen Huang, but the company has installed interim leadership and begun a successor search amid scrutiny of his conduct and alleged Chinese intelligence ties.
Intel disclosed the bandwidth target for its next Xeon server processors at SC26, alongside previously announced core and cache limits. Application performance, power requirements and launch timing remain unresolved.
Mercedes-Benz plans California early access to city-driving and hands-free highway assistance in December, ahead of a 2027 U.S. rollout. Mixed road tests and unannounced subscription prices leave buyers waiting for evidence of everyday value.
Researchers estimate that 403 large U.S. data centers consumed 81.8 TWh in a year. Their regional-grid analysis reveals substantial fossil-fuel dependence, but cannot isolate AI demand or fully expose the underlying facility data.
HSBC lowered its Zhipu share-price target to HK$500 and delayed forecast adjusted breakeven by two years. Improving API margins and substantial fundraising cushion a business whose gross profit still covers little of its research bill.