Qubit-Based Benchmarking of InP HEMT LNAs: Readout Fidelity Versus Power Consumption

AI summary

Overview

Research area: Quantum computing hardware — specifically cryogenic microwave electronics and superconducting qubit readout. The paper sits at the intersection of device physics (InP high-electron-mobility transistors), cryogenic engineering, and circuit quantum electrodynamics (cQED).

Technical level: Intermediate. The paper assumes familiarity with dispersive readout, IQ plane histograms, and noise temperature, but its central argument — that amplifier noise specifications do not predict qubit readout fidelity — is accessible to anyone with a graduate-level physics or electrical engineering background.

Scope: A single-author experimental study that benchmarks three state-of-the-art InP HEMT low-noise amplifiers (53%, 60%, and 70% channel indium content) plus a commercial reference against qubit single-shot readout fidelity, and maps that fidelity against dc power consumption from 75 µW to 10.5 mW.

What This Paper Is About

Cryogenic HEMT amplifiers at the 4 K stage are conventionally rated by their noise temperature, measured with a Y-factor technique using a matched resistive load — a number that says nothing about how the amplifier actually performs inside a qubit measurement. This paper asks a different question: if you put each amplifier into a real superconducting qubit readout chain and measure assignment fidelity directly from two-dimensional IQ histograms, do you get the same ranking that the noise temperature gives you? The answer is no, and the paper goes further by sweeping amplifier bias to show how fidelity trades off against dc power, which is a scarce resource on a dilution refrigerator.

Key Contributions

1. A qubit-in-the-loop benchmarking method. The two-dimensional IQ histogram of single-shot readout — specifically the empirically computed assignment fidelity, not an analytic Gaussian fit — is established as the primary figure of merit for cryogenic HEMT LNAs, complementing conventional Y-factor noise temperature characterization.

2. A head-to-head comparison of four amplifiers under identical qubit conditions. Three InP HEMT LNAs with 53%, 60%, and 70% channel indium content plus a commercial Low Noise Factory LNF-LNC4_8G reference are wired to four separate output chains and selected by a cryogenic switch at the mixing-chamber stage, so the qubit, input line, and room-temperature electronics are shared and only the amplifier under test changes.

3. A fidelity-versus-power map. For each device, assignment fidelity is measured across a drain-current sweep from 1.5 mA to 15 mA (75 µW to 10.5 mW of dc power), yielding the minimum power needed to reach 80% and 85% fidelity thresholds.

4. Identification of an unexplained high-power fidelity decline. Above roughly 5.5 mW, assignment fidelity falls for several devices, tracking an upturn in qubit thermal population that is not explained by the amplifier's reverse isolation or by 4 K-stage and mixing-chamber thermometry.

Main Findings

• Noise temperature does not predict readout fidelity. The 60% and 70% devices have similar Y-factor noise temperatures (1.1 K each) and both beat the 53% device (1.4 K), while the 70% device has the highest gain (45 dB versus 41 dB). Yet readout SNR and assignment fidelity peak at the 60% device: SNR 1.82 and fidelity 87.2%, versus 1.61 and 84.9% for the 70% device, 1.34 and 82.1% for the 53% device, and 1.57 and 83.8% for the commercial reference. The ranking is non-monotonic in indium content and matches the drain noise temperature reported in the prior Y-factor work rather than the overall noise temperature ranking.

• Fidelity does not saturate with power. Assignment fidelity rises steeply below about 1 mW, fluctuates around a plateau in the low-milliwatt range, and then declines above a device-dependent power around 5.5 mW. The decline tracks an increase in the qubit thermal population, defined as the fraction of shots prepared in |0> that are classified as |1>.

• Sub-milliwatt operation is sufficient for 80% fidelity. Across all four devices, fidelity above 80% is reached with as little as about 0.3 mW — the 70% device at 0.29 mW and the 60% device at 0.35 mW, roughly a factor of 22 to 27 below the 7.8 mW nominal bias. The 53% device needs 0.94 mW and the commercial reference 0.77 mW.

• The 85% threshold separates the devices more sharply. The 60% device needs 1.01 mW and the 70% device 0.77 mW to maintain fidelity above 85%, while the 53% device needs 3.73 mW and the commercial reference never reaches 85% anywhere in the tested range.

• The 60% device beats the commercial reference throughout the sweep. It exceeds the reference by an average of about 5% across the power sweep, with the largest statistically significant gaps (roughly 4% to 10%) in the 0.45 mW to 7.8 mW range.

• The 70% device has degraded since its original characterization. A repeat gain and noise measurement alongside this work, at drain-source voltage 0.7 V and drain current 15 mA, shows the 53% and 60% devices essentially unchanged (the 53% device moved from 41.87 dB and 1.63 K in 2024 to 41.99 dB and 1.65 K in 2026; the 60% device from 41.67 dB and 1.39 K to 41.39 dB and 1.44 K). The 70% device moved from 46.31 dB and 1.38 K to 45.58 dB and 1.67 K — a 0.73 dB gain loss and a 0.29 K noise temperature increase — so its true noise performance during the qubit measurements was somewhat worse than the table value.

• The high-power decline has no confirmed cause. Reverse isolation improves monotonically across the entire bias range tested, so it cannot explain the upturn, and the 4 K-stage and mixing-chamber temperatures do not track the increase in a consistent direction, ruling out simple self-heating of the shared cold stages. The author reports the effect as real and device-dependent rather than attributing it to a specific mechanism.

• Qubit parameters stayed stable across amplifier swaps (as a validity check). Energy relaxation time was stable to within 5% across all amplifier configurations, and any session in which a parameter drifted by more than 10% was excluded from the analysis.

Methodology in Plain English

The author built a readout line with a transmon qubit at the 10 mK stage of a dilution refrigerator, coupled to a coplanar waveguide resonator, with all the usual attenuation (70 dB from room temperature down to the sample). Rather than cooling down the fridge once per amplifier, four separate output chains were wired in parallel, each with its own HEMT LNA at the 4 K stage, and a cryogenic switch at the mixing-chamber stage chose which chain was connected to the qubit's readout line. Everything else — qubit, input line, room-temperature electronics — stayed identical.

For each amplifier, the qubit was prepared in |0> 200,000 times and in |1> 200,000 times, and each single-shot measurement produced one point in the IQ plane. Amplifier noise scatters these points into two blobs; the distance between blob centers carries the signal and the blob widths carry the noise. The fidelity was then computed empirically from how often a machine classifier assigned shots to the wrong blob, rather than assuming ideal Gaussian statistics — because the |1> blob is genuinely wider than the |0> blob, partly from qubit decay during the 3 µs integration window.

The power study repeated this at many bias points, with 10,000 shots per state at each point, sweeping drain current from 1.5 mA to 15 mA. Fidelity was then interpolated against dc power to find the minimum power needed for the 80% and 85% thresholds. The qubit was fully re-characterized before every amplifier swap or bias change to make sure the comparison was fair.

Why This Matters

Impact on research. The paper argues that a device specification measured in isolation can rank amplifiers differently from the metric that actually matters — qubit state discrimination error. It offers a reproducible, qubit-referenced protocol that any lab with a working qubit and a switchable output chain could adopt, and it flags a high-power fidelity degradation mechanism that Y-factor characterization structurally cannot see, since Y-factor uses a matched resistive load rather than a qubit coupled to a superconducting resonator.

Real-world applications:

• Power budgeting for large-scale quantum processors. Using a typical pulse-tube refrigerator's 1.5 W of cooling power at 4 K, an 85% fidelity requirement, and the best device, the paper estimates roughly 1500 simultaneous readout chains at minimum bias; with about 7 qubits per chain via frequency multiplexing, that could support up to roughly 10,000 qubits, versus only about 1000 qubits at the nominal bias of the least efficient device.

• Amplifier selection and bias-point choice in multi-qubit systems. The fidelity-versus-power curves give system designers a concrete basis for choosing an operating point instead of defaulting to the noise-optimized drain current.

• Procurement and qualification of cryogenic components. The qubit-based metric offers a way to qualify an amplifier against the measurement it will actually serve, not just its datasheet.

• Monitoring device aging in installed systems. The re-measurement finding that one device lost 0.73 dB of gain and gained 0.29 K of noise temperature over roughly two years suggests amplifiers in long-running systems may drift away from their qualification data.

Industry relevance. The results are directly actionable for companies building quantum processors and for cryogenic amplifier manufacturers: they suggest that the marginal fidelity return on amplifier power disappears well below conventional operating points, and that competing on noise temperature alone may not reflect real readout performance. The commercial reference amplifier is outperformed on fidelity by the 60% indium device at every power in the sweep.

Future Directions

• Identify the origin of the high-power fidelity decline. The paper explicitly leaves this open after ruling out reverse isolation and cold-stage thermometry. Understanding it would set a principled upper bound on useful drain current.

• Combine the best device with fast readout protocols. A prior demonstration with a commercial HEMT reached 99.5% two-state and 96.9% three-state assignment fidelity using a 140 ns readout, shelving to suppress decay error to a residual 0.03%, two-tone excitation, and machine-learning-assisted discrimination, without a quantum-limited amplifier. That work noted its fidelity could improve with a better amplifier, so pairing it with the 60% device is a natural next test.

• Extend the benchmarking to systems that include a JPA or TWPA. The paper frames the HEMT primarily as the dominant noise source when no quantum-limited amplifier is present, and as a secondary contributor — noise divided by parametric gain, plus a direct heat load at 4 K — when one is installed. Whether the same fidelity-versus-power trade-off holds in that regime is untested here.

• Track long-term device stability as standard practice. The 70% device's degradation was discovered only because the author repeated a 2024 gain and noise measurement in 2026, which raises the question of how often installed amplifiers should be re-qualified.

Target Audience

Experimental quantum computing groups building or scaling superconducting qubit systems; cryogenic and microwave engineers responsible for readout chains, amplifier bias, and 4 K-stage thermal budgets; amplifier manufacturers who want to know how their noise temperature specifications translate into qubit-level performance; and graduate students or postdocs setting up a dispersive readout measurement who need a practical method for comparing amplifiers within their own system.

Authors’ abstract

High-fidelity single-shot readout of superconducting qubits is essential for fault-tolerant quantum computation. Without a quantum-limited amplifier such as a Josephson parametric amplifier, the cryogenic high-electron-mobility-transistor (HEMT) low-noise amplifier (LNA) at the 4 K stage is the dominant noise source in the readout chain. HEMT LNAs are conventionally characterized by Y-factor measurements of the noise temperature $T_N$, independently of the quantum measurement they ultimately serve. Here we present a qubit-in-the-loop benchmarking method. The method uses the two-dimensional IQ histogram of single-shot readout as a direct figure of merit. We apply it to three state-of-the-art cryogenic InP HEMT LNAs with 53\%, 60\%, and 70\% channel indium content, together with a commercial reference amplifier. The 60\% and 70\% devices have similar $T_N$ in Y-factor measurements, both lower than the 53\% device. The 70\% device has the highest gain, but qubit readout SNR and assignment fidelity $\mathcal{F}_a$ instead peak at 60\%. Using this method, we map $\mathcal{F}_a$ against HEMT LNA dc power consumption for each device. $\mathcal{F}_a$ does not simply saturate with increasing power. For several devices it declines above a device-dependent power well below typical operating points. For the 60\% device, which has the highest overall $\mathcal{F}_a$, about 1~mW is needed to maintain $\mathcal{F}_a > 85\%$. Across all four devices, $\mathcal{F}_a > 80\%$ is reached with as little as about 0.3~mW, roughly an order of magnitude below typical operating points. These results establish a qubit-referenced benchmarking methodology and provide practical guidance for amplifier selection and bias in power-constrained multi-qubit readout systems.