In Google’s original paper, they noted that their Sycamore processor completed this task in 200 seconds and estimated it would take 10,000 years on the Summit supercomputer. IBM later provided a paper analysis indicating that with a different architecture that uses both RAM and hard drive space to store and manipulate the state vector, a high performance classical computer could do this in 2.5 days. Now two scientists from the Chinese Academy of Sciences have actually performed this calculation in a period of 5 days using 60 GPU cores. They generated one million correlated bit strings using Google’s circuit that had 53 qubits and 20 cycles and achieved a linear cross-entropy benchmark (XEB) fidelity of 0.739, which much higher than Google’s results. They are using a general tensor network method for the simulation which can be utilized for general problems and say that this method is much more efficient in computing a large number of correlated bit string amplitudes and probabilities than existing methods.

Compared to Google implementation this method using GPUs, is able to output the exact amplitude and probability of any bit string, it develops less noise, and they are able to compute conditional probabilities and sample accordingly. However, the full quantum approach still has significant advantages since it provided an answer in 200 seconds versus 5 days, a >2000X performance advantage. Also, the GPU would require exponential scaling when you start adding more qubits and more levels. So if Google were able to rerun their experiment with more qubits or more levels, the cost and/or the runtime of the GPU implementation would quickly make this approach intractable.

For more details on this experiment, you can view the paper posted on arXiv here.

March 5, 2021