200 seconds vs. 10,000 years

Google published its quantum supremacy paper in Nature on Wednesday. Its Sycamore processor, 53 working superconducting qubits, performed a specific computation in about 200 seconds that Google estimates would take the world’s fastest supercomputer about 10,000 years. IBM, which builds quantum computers too, published a response two days earlier arguing that with a better classical approach using lots of disk storage, the same task would take about two and a half days, not ten millennia.

I’m an engineer, not a physicist, so I spent some time trying to understand what the chip actually did.

The task is random circuit sampling. You apply a random sequence of quantum gates to the qubits, which creates a complicated quantum state, and then measure all 53 qubits, which gives you a 53-bit string. Repeat millions of times and you get samples from a probability distribution over all 2^53 possible strings, roughly 9 quadrillion. That distribution has a characteristic speckled pattern from quantum interference. Simulating it on a classical computer requires tracking the whole quantum state, which grows exponentially with the number of qubits. The quantum chip just runs it.

The chip is noisy, so the samples are only slightly closer to the ideal distribution than random guessing would be. Google verified this with a fidelity measure computed on smaller circuits that classical computers can still check, and extrapolated to the full size.

So is it supremacy? In the narrow sense of “a quantum device did a task that’s far beyond practical classical computation,” probably yes, even if IBM’s estimate is closer to right. Two and a half days on the largest supercomputer in the world versus 200 seconds is still a big gap, and adding a few qubits would make the classical simulation much harder again.

In the sense that matters to most people, whether it’s useful, no. Random circuit sampling was picked because it’s hard to simulate, not because anyone needs the output. The applications people talk about, like simulating molecules for chemistry, breaking RSA with Shor’s algorithm, or optimization, need either many more qubits or much lower error rates, and probably both. Breaking RSA is estimated to need millions of physical qubits with error correction. This chip has 53 noisy ones.

I think it’s a real physics milestone and an impressive piece of engineering, in the same category as the first flight at Kitty Hawk, which also didn’t carry passengers. I’d be surprised if quantum computers affect anything I work on in the next ten years. I’d be much less surprised if they matter enormously in chemistry and materials in twenty.