@qctrlHQ Key nuance: the 2e-5 error is after cultivation, keeping 25% of runs. After lattice surgery into the distance-5 surface code, fidelity stays above 0.99 ten rounds later at ~5% kept. The direct route also lets fast matching decoders handle it, a real plus for factories.
Q-CTRL on IBM's Nighthawk chip: a "magic state" with ~10x lower error than the prior record on Google's Willow (2 in 100,000), while keeping 3x more runs.
Magic states enable T gates for fault tolerance. Preprint (arXiv 2611.22035). @qctrlHQ@IBM#QuantumComputing
Q-CTRL on IBM's Nighthawk chip: a "magic state" with ~10x lower error than the prior record on Google's Willow (2 in 100,000), while keeping 3x more runs.
Magic states enable T gates for fault tolerance. Preprint (arXiv 2611.22035). @qctrlHQ@IBM#QuantumComputing
@Moth_Quantum@IBM A useful detail: each data qubit stores 2^m values (m = its address neighbors), but readout shots grow exponentially with m. So the real trade-off is shots vs data density. A square lattice like Nighthawk's allows m up to 4, i.e. 16 values per data qubit.
Moth Quantum loaded 1,568 numbers onto all 120 qubits of IBM's Nighthawk chip and read them back, vs 128 in the largest prior experiment of its kind.
No SWAP gates; images, audio, DNA. Preprint, not yet peer reviewed (arXiv 2611.18440). @Moth_Quantum@IBM#QuantumComputing
@IBMResearch@IBM@nvidia@Fermilab One reason this matters: good codes fail rarely, so estimating a logical error rate takes millions of simulated, decoded trials. Faster decoding for those studies (NVIDIA cites 12 weeks to 45 min) lets teams compare many more code and decoder designs before building hardware.
NVIDIA, with IBM + Fermilab: a new decoder cut one error-correction research computation from 12 weeks to 45 minutes (up to 13,000x faster, per NVIDIA).
It speeds up simulations used to test codes, not quantum hardware itself. @nvidia@IBMResearch@Fermilab#QuantumComputing
Qunova + IBM + RIKEN: an 80-qubit run on IBM Quantum System Two, paired with the Fugaku supercomputer, estimated an iron-sulfur cluster's energy within chemical accuracy of a top classical benchmark.
Extrapolated value; company release, Nov 12. @IBM@riken_en#QuantumComputing
@PsiQuantum@AMD A detail worth noting in the paper: the FPGA decoder takes the same number of clock cycles on every shot, so the machine never stalls waiting on an unusually hard one. Fixed latency is what makes feed-forward timing predictable, not just raw speed.
PsiQuantum + AMD: a transformer decoder on one AMD FPGA could decode a distance-8 error-correction block in 6.59 µs (10 µs target), finishing a block every 1.56 µs.
Fast decoding keeps error correction from stalling. Preprint, arXiv Nov 10. @PsiQuantum@AMD#QuantumComputing
@Cornell Easy to miss: in ion-trap machines, re-cooling the ions, not the gates, often sets the pace. Cooling all three measured motion modes at once with chip-delivered light attacks that overhead directly, and on-chip optics is also how many groups plan to scale to more ions.
Cornell cools a trapped-ion qubit's motion close to its quantum ground state in ~150 µs, vs ~500 µs with a standard laser setup, using light delivered by an on-chip photonic circuit.
Shown with one ion so far (PRX Quantum). @Cornell#QuantumComputing
@IBMResearch@IBM A circuit's size is qubits x operations. At 99.9% per gate, a 1,000-gate run is error-free only ~37% of the time, so "10,000 accurate ops" says more than the 520-qubit count. For the 1,040-qubit system, l-coupler fidelity (99.3% now, 99.9% goal) is the number to watch.
IBM unveils Nighthawk r3: a chip designed to run 10,000 quantum operations on 520 qubits (today's r2: 7,500+ on 120).
In 2027 IBM plans to link two r3 chips into a 1,040-qubit system. A design and a plan for now, not hardware yet. @IBMResearch#QuantumComputing
Quantinuum + RIKEN: Japan's Reimei quantum computer jumps from 20 to 56 qubits, all fully connected.
It now teams up with the Fugaku and ROQUO supercomputers on chemistry and materials projects; new results still to come. @QuantinuumQC@riken_en#QuantumComputing
Google DeepMind + Google Quantum AI: AI agents cut a quantum arcsin routine's cost 290x, with machine-checked proofs of correctness and cost.
Fusion sim: 10.9x cheaper, but still beyond early fault-tolerant machines. (arXiv) @GoogleDeepMind@GoogleQuantumAI#QuantumComputing
Quantum Motion shrank a spin-qubit readout sensor over 10,000x, to 18 square microns, by building a superinductor into a standard 22nm GlobalFoundries chip.
Signal-to-noise 4,500 in 4 ns. Not yet tested on a real qubit. (Nature Electronics) @quantum_motion#QuantumComputing
@psipbc Key scope: it bounds input reads, not total run time, for yes/no questions on most random inputs, so Shor-style periodic structure is untouched. Would love to see the polynomial spelled out: classical queries scale as roughly T to which power?
PSI: an AI system proved the Aaronson-Ambainis conjecture, open since 2008; Lean checked every step.
Upshot: on most random inputs, quantum computers can't cut input reads exponentially. Big speedups need structure, like Shor's. Not peer reviewed yet. @psipbc#QuantumComputing
@IBM On a sparse chain, a gate between distant qubits burns many SWAP hops into the depth budget. Four-neighbor couplers shrink that Manhattan distance so the same algorithm needs fewer layers before decoherence wins.