The Signal
QunaSys Simulates a 319 Million-Gate Chemistry Circuit
QunaSys says it simulated a 28-qubit, 319 million-gate quantum phase-estimation circuit for a four-hydrogen model using 1,024 NVIDIA H100 GPUs. The run took about 1.5 hours and produced an energy estimate within the stated chemical-accuracy
Quantum.world Team · Written with AI · fact-checked ·
QunaSys said on October 27 that it had simulated a 28-qubit, roughly 319 million-gate quantum circuit for estimating the ground-state energy of a four-hydrogen model. The calculation used 1,024 NVIDIA H100 GPUs on Japan’s ABCI-Q quantum-classical computing platform and took about 1.5 hours to generate roughly one million measurement results. The company says the most frequent energy estimate was 0.53 millihartree from the model’s exact solution, inside the 1.6-millihartree threshold it uses for chemical accuracy.
From 1.78 billion gates to 319 million
The circuit implements qubitization-based quantum phase estimation, or QPE, a method for extracting an energy value from a quantum representation of a molecule. QunaSys’s starting design was estimated at 99 qubits and about 1.78 billion gates. Circuit and state-preparation optimizations brought that configuration down to 28 qubits and about 319 million gates, making a full classical simulation possible on the GPU cluster.
The company used its QURI SDK Enterprise software to build and compile the circuit, then ran the simulation with NVIDIA cuQuantum. QunaSys says it distributed the quantum-state data across GPUs, reduced the amount of data exchanged between them and favored fast links among GPUs in the same compute node. It describes the result as the largest public end-to-end classical simulation of a qubitization-based QPE circuit it had identified as of 2026; that is the company’s qualified comparison, not a claim about every kind of quantum-circuit simulation.
One result tests accuracy; another tests size
The H₄ run checked more than whether a circuit could be executed. QunaSys analyzed both the resulting energy values and how frequently each appeared across the measurement results, testing the process from circuit construction through output analysis. The 0.53 mHa difference from the exact solution is below the 1.6 mHa chemical-accuracy benchmark cited in the release.
Separately, QunaSys expanded a hydrogen-chain model from two atoms to ten while allowing a larger energy-estimation error. Under those looser conditions, it simulated an H₁₀ QPE circuit with 42 qubits and about 490,000 gates. That result demonstrates a larger qubit count under different accuracy conditions; it is not the same calculation as the 319-million-gate H₄ run.
More GPUs brought a communication limit
A third test measured how GPU count affected runtime for a 32-qubit QPE circuit for H₂. In that fixed-workload test, processing time fell from about 4,155 seconds on one GPU to about 24 seconds on 512 GPUs, a reported speedup of roughly 173 times. Gains flattened beyond 512 GPUs as communication between processors added more overhead.
The distinction between the tests matters: the H₄ calculation demonstrates chemical accuracy for a small molecular model, the H₁₀ calculation probes circuit width with a relaxed error target, and the H₂ benchmark measures parallel performance. None was run on a quantum processor. The announcement presents classical simulation as a way to check that a future fault-tolerant algorithm’s full workflow produces expected results and to generate reference data for later comparisons—not as evidence of quantum speedup or practical molecular computation.
ABCI-Q supports a software check before hardware
The project used computing resources supplied through the National Institute of Advanced Industrial Science and Technology’s ABCI-Q Grand Challenge program. QunaSys also credited quri-parts-cuquantum, software developed with AIST under Japan’s Strategic Innovation Promotion Program, for multi-GPU simulation. The company plans to present the work at SC26 in Chicago, scheduled for November 15–20, 2026.
For now, the clearest result is a systems test: QunaSys linked circuit design, compilation, large-scale GPU simulation and measurement analysis, then compared the output with an exact energy for the same model. The headline circuit ran on 28 qubits; the separate 42-qubit result came with a looser error condition and a much smaller gate count.