Q Dev (@QuantumDevIO) header
Q Dev (@QuantumDevIO) profile picture
Nashville, TN
Joined November 2026
The @IBMResearch blueprint for quantum-centric supercomputing splits the classical side in two, per IBM: big CPU and GPU pools to clean up noisy results, and fast computers near the qubits to fix errors live. We think the second is harder to build. Which would you pick?
1
4
The updated @Alice__Bob roadmap aims for 200 error-corrected qubits built from 3,600 physical ones by 2030, then 2,048 from 89,600 by 2034, per alice-bob.com/roadmap. Neither machine exists yet. These are company goals. What has to go right for the 2030 date to hold?
1
7
Some quantum operations need special "magic" states. A @qctrlHQ team prepared one on an IBM Nighthawk chip that held 99.1% fidelity after ten stabilizer rounds in a distance-5 surface code, per arXiv 2611.22035. Only 4.6% of shots were kept. Can this work at higher distance?
1
16
Error correction eats qubits. Origin Quantum and USTC designed codes that keep over 1 usable qubit per 2 data qubits, about 6 times IBM's Gross-code rate at the same distances, per arXiv 2611.20410. Simulation only, on 5 or 6 stacked chip layers. Who builds that chip?
1
17
A GPU decoder from @nvidia, software that finds errors so they can be fixed, is aimed at ion trap and neutral atom machines, whose cycles last milliseconds. It averaged 0.64 to 4.97 ms per run at a 0.2% error rate on one GPU, per arXiv 2611.21321. Is that fast enough?
1
19
A code's distance is the fewest errors that can slip past it unnoticed. One 360-qubit code from @IBMResearch was listed as distance 24 or less because solvers timed out. A machine-checked proof now shows it is 24, per arXiv 2611.21236. Which code claims need proofs next?
1
14
A 250 MHz register-transfer-level simulation of a real-time decoder for quantum low-density parity-check codes gave a 173.8 ns mean and a 732 ns 99th-percentile latency, per arXiv 2611.19130, at a 0.5% physical error rate. Do we budget the 732 ns tail for the correction cycle?
1
15
On a 105-qubit Willow chip, @GoogleQuantumAI went after errors that pile up over time, per arXiv 2611.21578. A 3×3 walking circuit hit a logical error floor of <=3.20e-8 at 60 cycles, ~100x below the standard circuit. Radiation bursts set the limit. How do you shield for that?
1
11
Annealers often chain qubits into one logical spin. Chain-free maps on a @dwavequantum Advantage2 working graph reach 36×36 spins, per arXiv 2611.17842. At 19×19, layouts matched on mean energy still differed in magnetization and correlations. Is energy enough to judge samples?
1
6
A quantum data processing unit (QDPU) is timed at 12.5 µs to classify occupancy of a 50×50 array on one FPGA lane, per arXiv 2611.17652. NVQLink to an @nvidia DGX Spark had a 2.74 µs median round trip for 128-byte messages forwarded by a CPU core. What do we send to the GPU?
1
22
Qubit count gives a chip's size. University of Osaka measured percolating entanglement quality (PEQ) of 0.42 on a 64-qubit chip, per arXiv 2611.17941. Reanalyzing older @IBMResearch heavy-hex graph-state data gives 0.20 to 0.74. Should vendors report PEQ next to qubit count?
1
7
Quantum phase estimation usually needs controlled evolution. A team at MPQ ran it on an analog Rydberg simulator and read out absolute ground-state energies and gaps for up to 19 atoms in 1D and 2D, per arXiv 2611.18122. Which models would you point it at next?
1
11
Spin-adapted post-processing cut the diagonalized subspace for a transition-metal complex run on IBM hardware. Barison, Tacchino, Tavernelli, and coauthors at @IBMResearch apply it to sample-based quantum diagonalization (SqDRIFT) and study the spin energetics of the oxo-Mn(Salen) complex. The hardware runs were on the Heron r3 processor ibm_boston, with 3 Krylov times and 500 circuit realizations at 5, 10, and 15 excitations, per arXiv 2611.18071. At 15 excitations, the spin-complete construction reaches 0.84 mHa on the singlet, inside the paper's 1.6 mHa chemical-accuracy line, at a diagonalized dimension of 5.0 x 10^4. The Cartesian product needs 2.0 x 10^7 determinants to go further, a subspace about 400 times larger. The same pattern holds for the triplet and quintet, and the roots are exact spin eigenfunctions at every subspace size. The 4-aminobenzonitrile and FeN4 results are classical simulations. The hardware data are the oxo-Mn(Salen) runs on ibm_boston. When we budget a sample-based chemistry job, do we count the diagonalized subspace as carefully as circuit depth?
1
18
An autonomous correction protocol has pushed a photonic logical qubit past break-even. Ni, Xu, Zheng, and coauthors encode the logical qubit in a superconducting microwave cavity and protect it against photon loss with engineered dissipation, per Nature Physics. The protected logical qubit lasts about 2.0 ms, 18% longer than the best physical qubit in the same system. Measuring a small frequency shift, they report a 6.3 dB metrological gain relative to the most robust Fock-state superposition. This is an autonomous error-correction memory and a sensing demonstration, not a logical gate set. What error channel would we try this kind of recovery on next?
1
13
The first question we would ask is what speedup remains after error-correction overhead. Babbush, McClean, and coauthors conclude that quadratic speedups will not enable quantum advantage on early fault-tolerant devices unless error correction itself improves significantly, per their 2021 perspective. They used state-of-the-art surface-code constructions under a variety of assumptions. That conclusion still held if the logical-gate rate increased by more than an order of magnitude. Quartic speedups looked significantly more practical. This is a 2021 perspective based on surface-code cost models. When we pick a first fault-tolerant workload, what speedup degree are we counting on?
1
7
qLDPC codes get talked about for qubit overhead, but the constraint we would budget first is connectivity. Bravyi, Cross, Gambetta, and coauthors show a bivariate-bicycle memory that keeps 12 logical qubits for nearly a million syndrome cycles with 288 physical qubits at a 0.1% physical error rate, per Nature. The surface code would need nearly 3,000 physical qubits for the same performance. Connectivity is a degree-6 graph with two long-range edges per qubit that is not locally embeddable in a 2D grid. This is a circuit-level simulation of a fault-tolerant memory. If you were laying out a chip for this code, would you design around the long-range edges first, or the qubit count?
1
8
Fault-tolerant hardware is only useful if a classical decoder can keep up with the syndrome stream. Fall behind and the backlog stretches the run out exponentially. In Google Quantum AI's 2024 Nature experiment, a distance-5 surface-code memory stayed below threshold with real-time decoding, at an average decoder latency of 63 ± 17 µs over up to a million cycles, per Nature. Cycle time was 1.1 µs. They measure that latency from the last syndrome reaching the decoder software to a correction coming back. Feedback onto the qubits was not yet closed, and latency grows with code size. When you budget a logical cycle, how much time is left for the decoder and the return path?
1
12
Anderlini, Sestini, and coauthors trained and tested a variational classifier for beauty versus charm jets on Lagrange, a 5-qubit IQM Spark processor at Politecnico di Torino. Per the preprint, the data are simulated LHCb Run 2 open samples, not recorded collisions. A 4-qubit, 1-layer circuit, 100 training jets, and 256 shots reached a hardware AUC of 0.761, compatible with 0.763 ± 0.015 on a noiseless simulator, in 43 minutes of QPU time. This is a preprint of a small NISQ demonstration. No quantum-advantage result is reported. Which jet-tagging step would you want tried on a larger processor first?
1
50