How Researchers Use Nvidia’s GPUs to Simulate Qubits

Between integrating its Grace Hopper chip straight with a quantum processor and showing off the power to simulate quantum programs on classical supercomputers, Nvidia is making waves within the quantum computing world this month.

Nvidia is actually properly positioned to benefit from the latter. It makes GPUs that supercomputers use, the identical GPUs that AI builders crave. These similar GPUs are additionally worthwhile as instruments for simulating dozens of qubits on classical computer systems. New software program developments imply that researchers can now use increasingly supercomputing assets in lieu of actual quantum computer systems.

However simulating quantum programs is a uniquely demanding problem, and people calls for loom within the background.

Few quantum laptop simulations thus far have been in a position to entry multiple multi-GPU node and even only a single GPU. However Nvidia has made current behind-the-scenes advances, now making it attainable to ease these bottlenecks.

Classical computer systems serve two roles in simulating quantum {hardware}. For one, quantum laptop builders can use classical computation to test-run their designs. “Classical simulation is a basic side of understanding and design of quantum {hardware}, continuously serving as the one means to validate these quantum programs,” says Jinzhao Sun, a postdoctoral researcher at Imperial School London.

For one more, classical computer systems can run quantum algorithms in lieu of an precise quantum laptop. It’s this functionality that particularly pursuits researchers who work on purposes like molecular dynamics, protein folding, and the burgeoning discipline of quantum machine learning, all of which profit from quantum processing.

Classical simulations aren’t good replacements for the real quantum articles, however they continuously make appropriate facsimiles. The world solely has so many quantum computer systems, and classical simulations are simpler to entry. Classical simulations can even management the noise that plagues actual quantum processors and sometimes scuttles quantum runs. Classical simulations could also be slower than genuine quantum ones, however researchers nonetheless save time from needing fewer runs, in accordance with Shinjae Yoo, a computer-science and machine-learning workers researcher at Brookhaven Nationwide Laboratory, in Upton, N.Y.

The catch, then, is a dimension drawback. As a result of a qubit in a quantum system is entangled with each different qubit in that system, the calls for of precisely simulating that system scale exponentially. As a rule of thumb, each extra qubit doubles the quantity of classical reminiscence the simulation wants. Shifting from a single GPU to a whole eight-GPU node is a rise of three qubits.

Many researchers nonetheless dream of urgent as far up this exponential slope as they will handle. “If we’re doing, let’s say, molecular dynamics simulation, we would like a a lot larger variety of atoms and a bigger-scale simulation to have a extra real looking simulation,” Yoo says.

GPUs can pace up quantum simulations

GPUs are key footholds. Swapping in a GPU for a CPU, Yoo says, can pace up a simulation of a quantum system by an order of magnitude. That type of acceleration might not come as a shock, however few simulations have been in a position to take full benefit due to bottlenecks in sending data between GPUs. Consequently, most simulations have stayed throughout the confines of 1 multi-GPU node or perhaps a single GPU inside that node.

A number of behind-the-scenes advances at the moment are making it attainable to ease these bottlenecks. Nearer to the floor, Nvidia’s cuQuantum software program growth package makes it simpler for researchers to run quantum simulations throughout a number of GPUs. The place GPUs beforehand wanted to speak by way of CPU—creating an extra bottleneck—collective communications frameworks like Nvidia’s NCCL let customers conduct operations like memory-to-memory copy straight between nodes.

cuQuantum pairs with quantum computing instrument kits equivalent to Canadian startup Xanadu’s PennyLane. A stalwart within the quantum machine-learning neighborhood, PennyLane lets researchers play with strategies like PyTorch on quantum computer systems. Whereas PennyLane is designed to be used on actual quantum {hardware}, PennyLane’s builders particularly added the aptitude to run on a number of GPU nodes.

GPUs are key footholds. Swapping in a GPU for a CPU, Yoo says, can pace up a simulation of a quantum system by an order of magnitude.

On paper, these advances can permit classical computer systems to simulate round 36 qubits. In follow, simulations of that dimension demand too many node hours to be sensible. A extra real looking gold commonplace at this time is within the higher 20s. Nonetheless, that’s an extra 10 qubits over what researchers might simulate just some years in the past.

To wit, Yoo performs his work on the Perlmutter supercomputer, which is constructed from a number of thousand Nvidia A100 GPUs—sought for his or her prowess in coaching and operating AI fashions, even in China, the place their sale is restricted by U.S. authorities export controls. Fairly a number of different supercomputers within the West use A100s as their backbones.

Classical {hardware}’s position in qbit simulations

Can classical {hardware} proceed to develop in dimension? The problem is immense. The bounce from an Nvidia DGX with 160 gigabytes of GPU reminiscence to at least one with 320 GB of GPU reminiscence is a bounce of only one qubit. Jinzhao Solar believes that classical simulations trying to simulate greater than 100 qubits will possible fail.

Actual quantum {hardware}, at the very least on the floor, has already lengthy outstripped these qubit numbers. IBM, as an illustration, has steadily ramped up the variety of qubits in its personal general-purpose quantum processors into the hundreds, with bold plans to push these counts into the 1000’s.

That doesn’t imply simulation gained’t play a component in a thousand-qubit future. Classical computer systems can play necessary roles in simulating elements of bigger programs—validating their {hardware} or testing algorithms which may sooner or later run in full dimension. It seems that there’s rather a lot you are able to do with 29 qubits.