Quantum Supremacy, and a pathway to the practical future of quantum computing with machine learning
Last week, Google published a long-awaited paper as a cover story for Nature, announcing that it had achieved “Quantum Supremacy”, where Google’s quantum computer Sycamore solved a problem in 200 seconds that they claim would take the world’s fastest classical computer 10,000 years to solve.
The Quantum Supremacy experiment
The Supremacy experiment involved a sampling of so-called random quantum circuits, a task that is extremely difficult for classical computers, as demonstrated by leading computer scientist Scott Aaronson. Aaronson has long advocated for such an experiment and has worked closely with Google on its realisation.
At the same time, IBM posted a contradictory result in this blog post stating that they could simulate the Google experiment with a supercomputer in 2.5 days, not 10,000 years as estimated by Google.
The IBM result
IBM’s group demonstrated theoretically that using the Oak Ridge National Lab supercomputer Summit (currently the most powerful supercomputer on Earth with ~250 petabytes of hard disk space), one can simply store and manipulate the entire quantum state vector of Google’s 53-qubit Sycamore quantum chip in memory. Details of the full experiment as announced by IBM can be found here.
This is made possible by the size of Summit’s hard disk, which is the largest of its kind. The size of this hard disk allows storage of an amount of data that would be impossible to store by any other classical machine.
In doing so, IBM did in fact show theoretically (note that no experiments were actually run) that the whole quantum computation could be simulated in ~2.5 days rather than the 10,000 year estimate from Google. Google had based this estimate on a different algorithm called the Schrödinger-Feynman algorithm which is vastly slower, but requires much less memory than the algorithm IBM used.
Most important to note here is that a speedup of three minutes vs 2.5 days is a quantum speedup by a factor of 1,200 – a tremendous speedup by any measure.
As Aaronson detailed in his blog, a comparison in terms of elementary operations, i.e. using FLOPS (floating-point operations) as the equivalent of quantum gates, the quantum speedup becomes an even more impressive 40,000 billion times fewer operations than the world’s largest supercomputer.
Notably, increasing the number of qubits from 53 to 55 qubits would already be enough to make it impossible for even Summit’s 250PB storage to simulate the quantum system. A further increase by a few more qubits would require 10s or 100s of these supercomputers to simulate such a quantum system.
Overall, the message is clear: whether quantum supremacy has been technically achieved or not, we have undoubtedly seen the immense power of these machines.
The next big milestones for quantum computing
Moving forward from this notable moment, we think the next big milestones in quantum computing will be demonstrations of the practical usefulness of quantum computers.
On this path to useful practical applications, there has been a great deal of energy and advancement in the field of quantum machine learning. Recently there have been many results by a handful of teams across the world (including our team here at Rahko) that have significantly reduced the complexity in using quantum computers for chemical simulation.
One result we think is really cause for excitement belongs to one of our close collaborators, Giuseppe Carleo (paper here). Giuseppe and his collaborators have demonstrated a reduction by three orders of magnitude in the number of measurements required to estimate expectation values on a quantum computer using quantum machine learning. This was achieved by combining restricted Boltzmann machines with quantum circuits in a unique way.
The results allow us to perform quantum chemistry calculations three orders of magnitude faster than methods that do not use machine learning. This is another strong indication that we must rely on machine learning based techniques to enable practically useful near-term quantum computation.
Machine learning – the pathway forward to practical usefulness
Our team at Rahko has demonstrated in multiple instances that quantum machine learning is a far frontrunner in setting a path to practical, useful applications of near-term quantum computers.
Today, we are even more convinced of our approach and will continue to break ground to lay the path for the next important quantum computing milestones.
Leonard Wossnig, CEO of Rahko