Deep divefrom Episode 7AI & computing18 October 202623 min

Quantum Advantage on a Rentable IBM Chip

A deep dive into one paper: what was known before, what the researchers did, what the numbers say and what they don't.

The paper: Quantum computational advantage in random-circuit sampling on IBM superconducting quantum computers (Sedrakyan et al., BlueQubit, arXiv preprint)

MP3

Chapters

Also linked

Data and code: BlueQubitDev/rcs-nighthawk

Transcript

The script the two AI voices read from. The times are worked out from the script, not heard, so a line can start a few seconds early or late.

Read the whole episode

Cold open

Host 1You heard the trailer — this is the whole story. Welcome to the Hastrup Review of Science — where artificial intelligence reads real science, so you don't have to pretend you understood the paper. Today, a team says it ran a quantum advantage benchmark on a quantum computer anyone can rent through the cloud, with no special tuning.

Host 2The paper is called Quantum computational advantage in random-circuit sampling on IBM superconducting quantum computers. There are six authors, led by Tigran Sedrakyan, all at BlueQubit, a quantum software startup in San Francisco, with co-authors also at the École Polytechnique Fédérale de Lausanne and the XPRIZE Foundation. It is an arXiv preprint, version one, dated October sixteenth, twenty twenty-six.

Host 1So not peer reviewed, and not from IBM.

Host 2Correct on both counts. They used IBM's hardware through IBM's startup credits programme, and they state that their views are not IBM's. Where I add textbook background, I will say so.

Host 1The claim, in two sentences.

Host 2On a sixty-one-qubit patch of IBM's newest processor, they ran random circuits thirty-six cycles deep and collected a million samples in nineteen seconds, at an estimated fidelity of about two in a thousand. Their estimate of what a classical supercomputer would need to do the same job, under assumptions they chose to favour the classical side, is about a hundred and ten years on Frontier.

Why this field exists

Host 1Start with the basics. Why would anyone sample random circuits on purpose?

Host 2Background, not from this paper. A quantum computer with n qubits holds a state described by two to the n numbers, so at sixty-one qubits that is over two billion billion numbers. The hope is that some tasks become easy for the quantum machine and impossibly expensive for a classical one.

Host 1And random circuits are the cleanest test of that?

Host 2Background again. If you apply random gates over and over, the output becomes a probability distribution over bitstrings that has no shortcut structure for a classical computer to exploit. The quantum device just runs it and reads out bitstrings, while a classical computer has to calculate the probabilities, which for a deep enough circuit is believed to be exponentially hard.

Host 1Is it useful for anything?

Host 2Not directly, and that is widely accepted. It is a stress test, a deliberately hard task built to show where a quantum device stops being something a classical computer can imitate, not a calculation anybody needs.

Host 1How do you know the quantum machine got it right, if nobody can check?

Host 2That is the central awkwardness of the field. The standard score is linear cross-entropy benchmarking, usually called XEB: you check whether the measured bitstrings land more often on outcomes the ideal circuit makes likely. Averaged correctly, that score acts as a proxy for the circuit's fidelity, roughly the fraction of the time the machine ran the circuit without error.

Host 1And a fidelity of two in a thousand is enough?

Host 2Background. A noisy device mostly produces random junk, but a small signal of the true distribution survives, and the claim is that even that faint signal is classically expensive to fake. The fidelity you measure sets how much signal a classical rival would need to reproduce.

What was known before

Host 1What did we know before this paper?

Host 2The paper lists the history. Google's fifty-three-qubit Sycamore experiment came first, in twenty nineteen, then the Chinese Zuchongzhi processors, two point zero and two point one, then Sycamore at sixty-seven qubits, and Zuchongzhi three point zero with eighty-three active qubits at thirty-two cycles.

Host 1And the classical side fought back.

Host 2Yes, and the paper is open about that. Classical methods have cut the cost of Sycamore-class sampling by orders of magnitude; as well-known background, Google's twenty nineteen estimate of ten thousand years was challenged within days, and later work did that sampling task on classical hardware outright. Trapped-ion machines from Quantinuum, which can connect any qubit to any other, have meanwhile run random circuits at fifty-six and ninety-eight qubits.

Host 1So where does IBM fit in?

Host 2Background on the hardware. IBM's earlier chips used a heavy-hexagon layout, where each qubit talks to at most three neighbours, which is a poor fit for these benchmarks. The Nighthawk architecture changed that to a square lattice, which is the grid random-circuit experiments traditionally use.

Host 1What is new about Nighthawk, specifically?

Host 2According to the paper, it has a hundred and twenty qubits on a twelve-by-ten grid, with two hundred and eighteen tunable couplers linking every nearest-neighbour pair. This device is the second revision, called r two, named ibm phoenix; the authors had earlier run the same experiment on the first revision, ibm miami.

The open question

Host 1So what is the open question?

Host 2The field has learned that an advantage claim depends on details that are easy to blur. The paper names three crossovers: the depth where the output distribution spreads out evenly enough, called anticoncentration; the noise level past which small pieces of the chip dominate the score; and the depth past which the classical cost stops growing.

Host 1Take those one at a time. Why does the second one matter?

Host 2If the device is noisy enough, the score can be dominated by small subsystems that behave well on their own. Then a classical cheat can simulate those small pieces separately and match the score without doing anything hard, which is the spoofing problem.

Host 1And the third?

Host 2For a flat two-dimensional chip, the classical cost of computing one output probability grows with depth until it hits a ceiling set by the number of qubits. Beyond that depth, adding cycles mostly just lowers fidelity without making the classical job harder.

Host 1So the question for this paper is?

Host 2Can an off-the-shelf cloud processor, running the standard software stack, sit on the right side of all three crossovers at once, with a fidelity measured independently two ways. And can the authors say precisely which classical task their cost estimate refers to, because the paper stresses that different papers have priced different tasks.

What the paper did

Host 1Walk me through what they actually did. Start with the chip.

Host 2They aimed for an eight-by-eight block of qubits on the twelve-by-ten chip. There are fifteen ways to place such a block, and a program scored each one from the latest calibration data just before the run, using readout, single-qubit and two-qubit error rates.

Host 1And it threw out bad qubits.

Host 2Qubits with readout error above six percent were dropped, and couplers with two-qubit error above two percent were pruned. The winning placement kept sixty-one qubits, three were dropped, with a hundred and two couplers between them.

Host 1How good are the gates?

Host 2At the time of the run, the median two-qubit CZ gate error was about one point nine in a thousand, single-qubit gate error about two and a half in ten thousand, and readout error about eight in a thousand. So the error budget is dominated by the two-qubit layers and by readout.

Host 1What is in one cycle?

Host 2Every qubit gets an independent random single-qubit rotation, drawn uniformly over all possible rotations. Then CZ gates fire on one of four sets of couplers, horizontal even, horizontal odd, vertical even, vertical odd, in a fixed repeating order.

Host 1How many gates is that?

Host 2After pruning, those four sets hold twenty-nine, twenty-two, twenty-nine and twenty-two couplers, so about twenty-five and a half CZ gates per cycle. The thirty-six-cycle circuit has nine hundred and eighteen CZ gates, and the deepest benchmark at forty cycles has one thousand and twenty.

Host 1Now the hard part. How do they measure fidelity on a circuit nobody can simulate?

Host 2Two independent ways. The first is patching, borrowed from the Google experiments: cut the qubits into three or four patches and delete every CZ gate that crosses a patch boundary, so the circuit splits into separate blocks small enough to simulate exactly.

Host 1How small?

Host 2For three patches, about twenty, twenty and twenty-one qubits; for four patches, fifteen or sixteen each. The ideal probabilities for each patch were computed by exact simulation on ordinary processors on BlueQubit's platform, and the total fidelity is the product of the patch scores.

Host 1Doesn't one particular cut bias things?

Host 2That is why they randomise it. For each patch count they use five different partitions, three random circuits each, so fifteen circuits per data point, and the spread from the choice of cut goes into the error bars.

Host 1And there is a correction for shallow circuits.

Host 2Yes. Each patch score is divided by what a perfect device would score on the same patch, which removes an upward bias at depths where the outputs are not yet spread evenly; their patches start to anticoncentrate around twenty-four cycles.

Host 1What is the second method?

Host 2A mirror benchmark. You run a random circuit for half the depth, then its exact inverse, and check how often the qubits come back to the starting bitstring; that return probability is a fidelity estimate that needs no classical simulation at all.

Host 1That sounds almost too convenient.

Host 2It has known traps, and they address three. Systematic errors in the second half can partly undo those in the first, so they scramble each circuit with sixty-four random Pauli frames, which turns those coherent errors into random ones that add up instead of cancelling.

Host 1The second trap?

Host 2Readout on these chips is lopsided: a one is misread as zero about one point three percent of the time, a zero as one only about zero point two percent. Starting from all zeros would flatter the mirror, so each circuit starts from ten random bitstrings with half the qubits set to one.

Host 1And the third?

Host 2A plain mirror would contain more gates per cycle than the patched circuits, which lack their boundary gates. So the mirror skips each cycle's boundary gates for one of the five three-patch partitions, rotating through them, which matches the patched gate count within about two percent while keeping the whole chip connected.

Host 1How did they handle drift, the chip changing over the session?

Host 2They interleaved everything: at each depth, the mirror job and both patched jobs were submitted back to back, so slow calibration drift hits all curves alike. Readout was twirled on every job too, and the whole session used nine point one million shots and about eleven minutes of machine time.

Host 1And the headline run?

Host 2One million samples of the full, unpatched thirty-six-cycle circuit, which took nineteen seconds. The paper attributes the short session partly to a reduced delay between shots on r two; similar runs had taken several hours of machine time on r one.

Host 1Then the classical side. How do they price it?

Host 2They convert each circuit into a tensor network, a web of linked arrays, and search for the cheapest order in which to contract it into one output probability, using the open-source tools quimb and cotengra. The cost of the best order found is counted in complex operations, and they state their search budgets, because a short search can overestimate the cost by two to five orders of magnitude.

Host 1How do they know their estimator is honest?

Host 2They ran it on the five published Google tensor networks and matched the published costs to within zero point seven orders of magnitude on all five, and to within zero point zero one on Sycamore at sixty-seven qubits and thirty-two cycles, the case closest in depth to theirs. On a reconstructed Zuchongzhi three point zero network they landed within a factor of four of that team's estimate.

A word from our sponsor

Host 1Before the results, a word from our sponsor.

Host 1Is your laundry drifting out of position? Are your shirts losing coherence on the line? The Hastrup Review of Science is brought to you by clothes pegs.

Host 2I have used the same set for years. They hold their state through wind, rain and the occasional neighbour.

Host 1A clothes peg applies a simple two-point clamp, with no calibration, no cloud account and no error bars. Each one works independently, so you can add as many as your line can hold.

Host 2Pin from the corners. It is always better to secure the edges before the middle starts to sag.

Host 1Clothes pegs are reasonably priced and in stock right now. Our referral link is in the show notes.

Results

Host 1So, the results. Do the two fidelity methods agree?

Host 2Yes. The mirror curve runs from four to forty cycles, the three- and four-patch curves from twenty to forty, and at every depth where they overlap they agree with each other across more than two orders of magnitude of decay.

Host 1What is the decay rate?

Host 2A single exponential fit to the mirror data gives a fidelity of zero point eight seven two per cycle. On the first revision, the same fit gave zero point eight three six.

Host 1That sounds like a small difference.

Host 2Per cycle it is, but it compounds. Over thirty-two to forty cycles it adds up to more than a factor of ten in fidelity, which the authors call a large step between the two chip revisions.

Host 1And the number at thirty-six cycles?

Host 2The fit gives two point three in a thousand; the directly measured patched and mirror values range from one point eight to two point two in a thousand. Importantly, those come from proxies measured at the same depth, not from extrapolating a shallower curve.

Host 1Why thirty-six cycles, and not deeper?

Host 2Because of the classical cost curve. Below the ceiling, the cost of one amplitude grows by about three hundred times every four cycles; at thirty-six cycles it reaches roughly ten to the twenty-two complex operations, and from thirty-six to forty it changes by less than a factor of two.

Host 1Ten to the twenty-two. Put that in human terms.

Host 2The paper says that one single output probability would take about two days of the entire Frontier supercomputer. That is also why nobody can check the million samples directly: you would need about a million of those amplitudes.

Host 1And the ceiling lands where theory said it would?

Host 2Roughly. A rule of thumb puts the ceiling near four times the square root of the qubit count, about thirty-one cycles here, and they measure it between thirty-two and thirty-six. The four, rather than the three seen in earlier experiments, comes from the CZ gate being a weaker entangler than the gates Google and the Zuchongzhi team used, so it takes longer to reach the ceiling but the ceiling itself is the same.

Host 1Now the hundred and ten years.

Host 2The classical task they price is to draw a million samples matching the experiment's fidelity, using a known method called bounded-fidelity rejection sampling, which needs about ten amplitudes per sample and gets cheaper in proportion to the fidelity. At thirty-six cycles that comes to one point two times ten to the twenty-seven floating-point operations, about a hundred and ten years on Frontier running at twenty percent of its peak.

Host 1And at the other depths?

Host 2At forty cycles about a hundred years, because the cost has stopped growing while the fidelity keeps falling. At thirty-two cycles only about two years, which is why thirty-six is the sweet spot where fidelity times cost is largest.

Host 1What about cheating, the spoofing you mentioned?

Host 2They tested it on their own circuits. The obvious cheat is a severing attack: cut the chip where it is weakest, simulate the pieces separately, and hope the score holds up.

Host 1Where is it weakest?

Host 2Because the placement looked only at error rates, the coupler graph has weak spots: a balanced cut through just six couplers, an eleven-qubit block attached by five, small arms attached by three, and eight qubits hanging on only two couplers. Those are exactly the cuts a cheater would pick.

Host 1And?

Host 2None of them reaches the experimental score at thirty-six cycles. The best, an arm mode, scores about two point two in ten thousand, with an uncertainty almost as large, so it is a factor of ten short; the wider cuts are six or more orders of magnitude short.

Host 1Was that also true on the first chip?

Host 2No, and that is a nice detail. On r one, some of those small-arm cheats did reach the target score and had to be ruled out by arguing they were too expensive to run; on r two the higher fidelity raises the bar, and the cheats simply fall short.

Host 1And the even-spread condition, anticoncentration?

Host 2By their check, the outputs are essentially fully spread by thirty-two cycles and stay that way at thirty-six and forty. So the anchor depth is well clear of the shallow regime where known spoofing tricks work.

Host 1How does it compare with earlier records?

Host 2They put every major experiment under one yardstick, and this is where modesty is called for. By that measure the result sits above Sycamore at fifty-three qubits and the early Zuchongzhi runs, but well below Sycamore at sixty-seven qubits, at about fifteen hundred years, Zuchongzhi three point zero, at around two hundred million years, and Quantinuum's ninety-eight-qubit Helios, at around three hundred trillion years.

Host 1So it is not a record.

Host 2Not by raw classical cost, and the paper does not claim one. The claim is narrower: the first advantage-scale random-circuit sampling on a commercially and broadly accessible processor, using the standard cloud stack, that non-experts could replicate.

Why it matters

Host 1Step back. What is the principle here?

Host 2A quantum advantage benchmark is not a property of a chip, it is a position: depth against the classical cost ceiling, noise against the spoofing threshold, and a specific named classical task. This paper is careful to state all three, which earlier claims did not always do.

Host 1What changes, and for whom?

Host 2For researchers, the circuits, layouts, patch partitions and the million measured samples are released openly, so classical-simulation groups have a concrete target to attack and others can rerun it on the same public hardware. For people building these experiments, there is a practical design rule.

Host 1Which is?

Host 2When choosing qubits, do not optimise gate errors alone; also avoid placements with weakly connected corners, because those are where spoofing attacks get in. They argue this would widen the margin at essentially no cost in fidelity.

Host 1And for IBM?

Host 2It is a third-party measurement showing the second Nighthawk revision is substantially better than the first on a demanding benchmark, more than ten times higher fidelity at thirty-two to forty cycles. That is an independent data point, although the authors received IBM credits to do it.

Host 1What does it not mean?

Host 2It does not mean quantum computers can now do anything useful that classical ones cannot; random-circuit sampling has no known application. It does not mean a hundred and ten years is a firm lower bound, and it does not mean the sampling has been verified directly, because it cannot be.

Caveats

Host 1Now the caveats.

Host 2First, it is a preprint from a company that sells quantum software, so it has not been through peer review and the authors have an obvious interest in the result. The data and code release does make it unusually checkable.

Host 1Second?

Host 2The hundred and ten years is an estimate for one specific algorithm, single-amplitude contraction with the best paths they found. The authors list what it leaves out: reusing work across many amplitudes, approximate contraction and matrix-product-state methods could all lower the cost, and they write that the history of this benchmark says some will.

Host 1But they assumed unlimited memory, which helps the classical side.

Host 2Yes, and it is a big assumption. Real memory limits can raise the cost by up to twelve orders of magnitude on published networks, so by quoting the unlimited-memory floor they pick the conservative direction for their claim, but the number is also very sensitive to whichever memory model a critic chooses.

Host 1Third?

Host 2The fidelity at thirty-six cycles comes from proxies, patched circuits and mirrored, slightly modified circuits, not from the actual circuit that was sampled. The two methods agree, which is reassuring, but the XEB score does not certify that the samples are close to the true distribution in a strict sense.

Host 1Fourth?

Host 2The spoofing analysis covers severing attacks, which is the family they consider most relevant, and the paper calls its conclusions relative to that attack class. Other noise-based classical algorithms exist, and the paper acknowledges that under specified noise models, noisy random circuits can be sampled classically outright.

Host 1And scaling?

Host 2They state it themselves: on a flat lattice at fixed error per gate, the cost of cutting the chip grows only like the square root of the qubit count, while total error grows with the qubit count. So simply making the chip bigger without making gates better eventually hands the advantage back to the spoofers.

Host 1Anything on the comparison table?

Host 2Yes. The yardstick is deliberately uniform, not the cheapest known attack on each row; for the first Sycamore circuit, the paper notes that a batched classical method reduces the sampling to six seconds on Frontier. So the table ranks difficulty under one convention, not real-world classical cost.

What to watch next

Host 1What should we watch for?

Host 2Classical rebuttals, almost certainly. With the circuits and samples public, tensor-network and approximate-simulation groups can try to bring the hundred and ten years down, and that is how earlier claims of this kind were tested.

Host 1And on the quantum side?

Host 2Whether outside groups rerun it on the same public device using the released code and get the same fidelity curves, which would test the replicability claim directly. Also whether future placements follow the design rule and avoid weakly connected corners.

Host 1Anything longer term?

Host 2Whether fidelity improvements keep pace with chip size, given their own scaling caveat. The jump from zero point eight three six to zero point eight seven two per cycle is the kind of progress that pushes the spoofing crossover outward; the open question is whether it continues.

Sign-off

Host 1That is the deep dive: Quantum computational advantage in random-circuit sampling on IBM superconducting quantum computers, by Sedrakyan and colleagues at BlueQubit. The link is in the show notes.

Host 2Short version: sixty-one qubits on a publicly rentable IBM chip, a million samples in nineteen seconds at a fidelity two independent methods agree on, and an estimated hundred and ten Frontier years under classical-friendly assumptions. It is a preprint, the classical estimate covers one algorithm, and the open data invites the counterattack.

Host 1Every Friday this same feed carries the roundup episode: five new papers as trailers, and you pick which become deep dives. This has been the Hastrup Review of Science.