A team from Oak Ridge National Laboratory, NASA, University of Illinois, and Google recently posted a paper showing simulation results that were run in Summit, the world’s fastest supercomputer located in Oak Ridge, Tennessee. Summit consists of 9216 Power9 CPU’s, 27,648 NVIDIA GPUs, 600 GB of DRAM memory supplemented by 800GB of non-volatile memory on each of the 4608 nodes and consumes almost 14MW of power at peak utilization. It cost $325 million and was brought on-line in June 2018.
Part of the reason for this exercise was to establish a methodology for determining quantum supremacy and to set the standard for the largest circuit a quantum computer needs to surpass in order to declare quantum supremacy against the best classical computer.
The circuit that was chosen as the standard benchmark for the experiment is called the Random Quantum Circuit (RQC) and the simulation program chosen is called qFlex, a higher performance RQC simulator based upon a tensor contraction algorithm that was developed by Google and NASA. For the experiment, qFlex was redesigned and reimplemented to run on the GPU based Summit computer as efficiently as possible.
In the paper, the team showed results of a simulation of a 7 x 7 qubit grid (49 qubits) with a gate depth of 40 (not including beginning and ending Hadamard gates) that took 2.44 hours of runtime and consumed about 21.1 MWh of electricity. They also showed results of simulating a 11 x 11 qubit grid (121 qubits) with a gate depth of 24 that took 0.278 hours of runtime and consumed about 2.32 MWh of energy.
One thing to remember is that runtime on this machine is very expensive. Just the electricity alone for these runs likely cost several thousands of dollars. One point made in the paper is the relative power efficiency of a quantum computer. They point out that a 72 qubit machine based upon Google’s Bristlecone chip will consume about 15 KW of power with the majority of that used by the dilution refrigerator. In comparison, the Summit computer needed about 8 MW of power while the calculation was running.