One thing not fully understood is that significant innovation is still occurring in classical computing and those trying to demonstrate a quantum advantage over a classical solution will find that the competitive bar is continually being raised. Some will say that classical computing is slowing down because Moore’s Law improvements are becoming harder to come by, but that is too simplistic. There is still much classical computing innovation still occurring in new architectures, new algorithms, and development of quantum-inspired solutions. Two recent announcements from GPU manufacturer NVIDIA and a research group in China show some significant advances in the classical simulation of quantum algorithms that may make these approaches much more attractive for use in real world problems.
The first announcement has come from NVIDIA which announced a software product called cuQuantum earlier this year that was specificially designed to accelerate quantum simulations. This product contains two libraries. The first is called cuStateVec which allows for a state vector simulation that supports tens of qubits, depending upon how much memory is available. The second is called cuTensorNet which uses a tensor network simulation approach that can potentially simulate some algorithm that require thousands of qubits. The news this week is that cuStateVec has now progressed to a public beta status and the cuTensorNet is expected to reach this state in December. Previously, this software was still under development and internal testing at NVIDIA. In addition, cuStateVec has already been integrated into qsim, Google Quantum AI’s state vector simulator and is slated to be integrated next month into Qiskit AER, a simulation framework from IBM.
To show the potential power of these simulators, NVIDIA used their cuTensorNet library with their Selene supercomputer that has 896 GPUs to solve a Maxcut problem with 3,375 vertices. This required simulation of a quantum circuit that had 1688 qubits, an 8 times improvement over previous attempts. For more about NVIDIA’s activities in using GPUs for simulation of quantum algorithms, can you view our previous article from earlier this year when they first announced it and also their latest announcement that describes their public beta, partnerships with IBM, Google, IonQ, Pasqal and others and also their large MaxCut demonstration.
You may remember that when Google announced their “quantum supremacy” experiment (which they have since renamed to “beyond classical”) they indicated that finding a classical solution to the problem would take the largest classical computer about 10,000 years compared with 200 seconds on their Sycamore processor. Very shortly after that, IBM posted a rebuttal and provided a paper analysis of a different simulation algorithm and felt that the classical simulation could be accomplished in about 2.5 days. Now a group from the Chinese Academy of Sciences and Peking University has implemented an approach that achieves this in 15 hours using a cluster of 512 GPUs. They used a tensor network approach to solve the uncorrelated sampling problem based on contractions of the three-dimensional tensor network. Furthermore, they indicated that the times an be further improved by either by using the cuQuantum libraries from NVIDIA or using a modern supercomputer that can provide exaflops performance. Overall, they estimated that the time to solution could be reduced to a few dozens of seconds which would be much faster than Google’s result on Sycamore. So the quantum supremacy would not supreme anymore! For more information about this team’s approach to speed up the simulation times, you can view an arXiv preprint that they recently posted here.
November 9, 2021