On July 29, 2026, HRL Laboratories published results in 'Nature' showing that it integrated a cryogenic CMOS controller and silicon qubits within a single dilution refrigerator, running a repetition code without receiving real-time commands from room-temperature equipment. The device can configure up to 18 qubits from 54 quantum dots, and for the distance-5 repetition code, the logical error rate was 4.7 times lower than that of a distance-3 subsystem extracted from the same data.
HRL describes this device as a "self-driving quantum processor." In reality, it is not a machine that devises quantum algorithms on its own. What is novel is that a once-written instruction sequence is executed by a 4K control chip, removing the configuration in which a room-temperature waveform generator continuously sends pulses one by one. Before scaling up the number of qubits, this research answers the problem of ballooning control wiring and heat generation using semiconductor manufacturing techniques.
The 4K Controller Removes the Room-Temperature Rack from the Real-Time Path
HRL's quantum processing unit (QPU) consists of a control board operating at 4K, a qubit substrate in the millikelvin regime, and superconducting ribbons connecting the two. The qubits are "exchange-only" type, encoding information in the combined spin of three electrons confined in a Si/SiGe quantum well. Instead of using microwaves to induce resonance, the exchange interaction between quantum dots is switched using voltage pulses, making it easier to pair with baseband signals generated by cryogenic CMOS.
The control chip is fabricated using commercial 130nm RF CMOS and carries approximately 70 million transistors. Multiple sequencers operating at up to 250MHz read user programs from a 6,144-word instruction memory. Once a program is initialized, it can process instructions without additional communication with room-temperature equipment. Power consumption during typical operation is below 3.5W.
The analog section features 156 outputs, 366 DACs, and 78 pulse generators. It generates 150 time-varying control signals on the 4K side, sending them to the qubits along with 35 static biases. Pulse widths can be set from 400 picoseconds to 6 nanoseconds. The design shifted from routing precise analog waveforms from room temperature to each individual qubit, to instead passing digital instructions to the cryogenic side.
From the perspective of qubits with an average electron temperature of 150mK, the 4K chip becomes a significant heat source. To address this, HRL arranged 296 niobium wires on approximately 1cm-wide polyimide. This suppresses adjacent crosstalk below -80dB up to 10GHz, while keeping heat transferred from the 4K stage to the mixing chamber below 10 microwatts. The round-trip signal delay is approximately 3 nanoseconds. Simply moving the controller to cryogenic temperatures would not have been sufficient—the key was co-designing high-density wiring that resists heat transfer alongside the controller itself.
Running 200 Rounds at an Average CNOT Error Rate of 0.35%
The peer-reviewed published version shows average gate error rates of 0.017% for single-qubit operations and 0.35% for CNOT. The minimum reproducible CNOT error rate was 0.09%. The CNOT arranges 37 exchange pulses across 45 timesteps, with pulse intervals adjusted between 12 and 28 nanoseconds. Since HRL's 2023 two-qubit experiment achieved an encoded CNOT fidelity of 96.3%, today's average value is consistent with the paper's assessment of roughly an order-of-magnitude improvement in error rate.
Not all of this improvement can be attributed to the quantum dot material itself. Intrinsic charge noise was reduced to less than one-tenth of that in previous HRL devices, and the combined contribution of charge noise and magnetic noise to CNOT error was estimated at only 0.02 percentage points. What dominated the observed errors were static magnetic field gradients and context-dependent deviations in pulse waveforms. In spin-level simulations, calibration-drift terms positioned to match the measured CNOT errors accounted for approximately 80% of detection events. This is a model-based attribution of causes, not a directly measured proportion.
For the distance-5 repetition code, 5 data qubits and 2 ancilla qubits were used. Syndrome extraction was varied from 1 to 200 rounds, with 50,000 shots repeated at each point. A single instance of 200 rounds alone involved 1,805 initializations, 805 measurements, and 335,422 exchange pulses. The distance-5 logical error rate was 0.50%, while the average for distance-3 subsystems within the same experimental data was 2.4%, yielding an improvement factor of 4.7 when increasing the code distance. However, unexplained variations in error rate remained in the distance-5 data.
In a separate 6-physical-qubit experiment, a [[4,2,2]] error-detection code embedding 2 logical qubits into 4 data qubits was run for 3 rounds. When approximately 77% of the 10,000 shots for each input state were discarded based on detection results, the fidelity of the 2-logical-qubit state reached 0.95. The average ignoring detection results was 0.59. This high value of 0.95 is not the performance after correcting and retaining all shots—it is the value after excluding trials where errors were detected.
"Self-Driving" Refers to Instruction Execution and Some Automated Calibration
What makes the cryogenic controller autonomous is the part where it interprets written instructions at each clock cycle and updates analog output settings. Qubit readout signals are still digitized at room temperature today, and static current/voltage biases are also supplied by the room-temperature system. Both manual and automated methods were used to determine the biases applied to electrodes. Therefore, this is not a device with all electronics outside the refrigerator removed.
The calibration software "Quiver" includes automated routines for loading electrons into each quantum dot and tuning tunnel coupling. A convolutional neural network that reads charge stability diagrams was trained on thousands of human-labeled data points. The supplementary materials also show a mechanism for recording, in tree structure, the sequence of operations from fully automated single-qubit calibration, allowing later tracking of the success or failure of unattended operation spanning hours to days. Measuring multiple locations in parallel accelerated the electron-loading process by 4 to 5 times.
Even so, the device does not autonomously handle arbitrary failures on its own. Due to variations in materials and manufacturing, calibration is exploratory, and humans have built in processes to return from dead ends to known states. What is called "self-driving" here is automated operation combining a deterministic cryogenic sequencer with calibration software that includes conditional branching. This is not a demonstration of general AI replacing researchers.
Beyond 18 Qubits, Walls of Wiring Manufacturing and Power Remain
HRL demonstrated universal two-qubit gates using 6 quantum dots in 2023, and in 2025 reported single-qubit fidelities exceeding 99.9% using a scalable two-dimensional array with three-layer wiring. The current chip arranges 54 quantum dots in three columns, capable of configuring up to 18 qubits into a 3×6 grid. The advance here is unifying logical gates, two-dimensional wiring, and cryogenic control into a single device on which error-correction protocols can run.
However, the distance-5 experiment used 7 physical qubits, and the [[4,2,2]] experiment used 6 physical qubits—the full 18 qubits were not error-corrected as a whole. The types of errors protected against by the repetition code are also limited, and the [[4,2,2]] scheme works by detecting errors and discarding trials. The paper itself lists as unresolved items the interconnects and back-end qubit-side wiring needed for transition to mass production processes, control power per qubit, and device uniformity. A system-wide error-correction design tailored to exchange-only qubits will also be necessary.
Six days before the announcement, IBM finalized a definitive agreement to acquire HRL from Boeing and General Motors. In its acquisition announcement, IBM cited cryogenic technology and control electronics in addition to spin qubits. Interconnects and packaging are also included. Today's QPU serves as evidence that this surrounding technology has reached the point of operating as a single integrated device. However, the acquisition is subject to regulatory approval and other conditions, with expected completion by the end of Q3 2026. How HRL's research results will be incorporated into IBM's product plans will need to be confirmed through the roadmap after the deal closes.