The world has gone AI-crazy—or so it may seem, with the quest for new AI models developed by OpenAI, Mid Journey, Anthtropic, Gemini, or one of the many AI companies now focused on building AI models and services. What hasn’t gone unnoticed is the reliance on a specialist type of hardware called a GPU. That GPU is made primarily by one company, NVIDIA.
The GPU, or Graphical Processing Unit, is at the heart of the training and inference of these revolutionary AI models. Without the specialist hardware device, developing or building models that are becoming embedded in our daily usage, such as chatGPT, is almost impossible.
But could there be another play by NVIDIA (NASDAQ: NVDA)? There is a connection between NVIDIA and Quantum Computing. Let me explain. When a quantum programmer wants to simulate a quantum computer, they need a supercomputer. This might seem somewhat counterintuitive. However, most of the time, developers of quantum algorithms will simulate their quantum circuits on a traditional, classical, or [conventional] computer.
The point is, of course, to run those output circuits on a real quantum computer with ‘real’ physical qubits (qubits are quantum bits), which have quantum mechanical properties that can be exploited to create effects that cannot be replicated on conventional machines. However, for many purposes, the array of qubits offered by hardware makers in reality has not been that vast and is difficult to simulate. You could simulate a small quantum computer on various devices, some reasonably primitive, because it is not until a middling tens of qubits that simulation becomes troublesome. If the circuit size is small enough, you can run a quantum simulation on a Spectrum ZX from the 1980s – don’t expect rapid results in real time!
Naturally, circuit sizes and the number of qubits have increased, as well as the physical capabilities to build those qubits into quantum computers of larger and larger scale. Some makers of quantum computers are “shipping” quantum computers with over 1,000 qubits. IBM Condor has its 1,121 Qubit processor, which rocks an impressive number qubits. Not only is the start-of-art today impressive, but the continued growth of the qubit count towards ever more impressive numbers, even by IBM’s aggressive roadmap.
In the age of hardware acceleration, GPUs, as you’d expect, are finding their way into many applications. Therefore, it comes as no surprise that they are deployed in the simulation of quantum circuits. NVIDIA, one of the stock market darlings of the moment, produces a range of hardware, and companies are clamoring to buy as many GPUs as they can make.
NVIDIA Corporation, founded in 1993 by Jensen Huang, Chris Malachowsky, and Curtis Priem, has evolved from a graphics processing unit (GPU) pioneer to a global leader in visual computing and artificial intelligence (AI). Initially focusing on GPU technologies for gaming and multimedia applications, NVIDIA expanded its scope to encompass high-performance computing (HPC), data centers, autonomous vehicles, and more. The company’s CUDA parallel computing platform and Tesla GPU accelerators revolutionized scientific computing, enabling breakthroughs in fields such as deep learning.
CPUs vs. GPUs vs. TPUs vs. QPUs
Those three-letter acronyms are cropping up everywhere. Likely, you’re using a CPU every day of your lives. But GPUs are becoming more mainstream; once the preserve of severe gamers for acceleration, these graphical processing units were used to speed up the display of graphics, which consisted of the same workload that lent itself to acceleration with hardware.
TPUs are Tensor Processing units, and the moniker comes from Google’s stable. QPUs are quantum processing units and represent putting “Quantum” in processing units, something fundamentally different. QPUs, no matter from the hardware manufacturer, typically exploit the quantum physics of superposition and entanglement.
The Rise of Hybrid Quantum and Classical Computing
As the world continues to explore quantum computing, real world cases are emerging especially in the classical-hybrid quantum computing space. This is where users run circuits that contain classical and quantum elements. You might argue that this is standard case anyway, however, as no machines are fully quantum and classical information of course still needs to be quantum encoded. Developers are hooking up classical computations such as neural networks with quantum components in so called hybrid worflows. Ditto for Drug Discovery. The question that many have, is what benefit the quantum components have on workflows and it may be more complicated than whether or not there is an end-to-end speed-up.
NVIDIA Quantum
cuQuantum Appliance helps developers get started by making simulation software available in a container optimized to run on the latest NVIDIA DGX and HGX systems. The stack includes Google’s Cirq framework, qsim simulator, and NVIDIA cuQuantum.
The appliance software is claimed to have achieved best-in-class performance on critical problems in quantum computing, including Shor’s algorithm, random quantum circuits, and quantum Fourier transform. Recent software updates to our container offering have enabled a 4.4X speedup over previously reported numbers. Combined with ~2x speedups offered by Hopper GPUs, users see even more significant speedups over CPU implementations despite CPU hardware and software improvements.
NVIDIA CUDA-Q
NVIDIA CUDA-Q is an open-source programming model designed for building quantum-classical applications, allowing the utilization of heterogeneous computing architectures like quantum processing units (QPUs), GPUs, and CPUs simultaneously to address real-world problems. By offering tools to program these architectures harmoniously, CUDA-Q enables the acceleration of quantum applications, crucially supporting the scalability necessary for quantum computing workloads.
Circuit simulation for quantum systems is often limited by the immense memory requirements to store the state vector. NVIDIA’s Hopper H100 GPUs, with their 80 GB of memory each, provide a solution beyond the constraints of current QPU hardware. The CUDA-Q framework, alongside the nvidia-mgpu target, effectively pools the memory of multiple GPUs within a node and across clusters, facilitating scalability and removing the memory bottleneck associated with single GPUs. This enables the manipulation of quantum state vectors for circuit simulation, with memory usage scaling according to the available GPU resources.
The multi-QPU mode within CUDA-Q allows for the parallel programming of quantum workflows, significantly reducing runtime by leveraging the available compute resources. For instance, tasks such as circuit cutting protocols, which traditionally execute subcircuits sequentially, can now run in parallel, substantially decreasing runtime. By defining multiple endpoints, each capable of simulating independent parts of a problem, CUDA-Q’s multi-QPU mode enables asynchronous parallel execution, as exemplified by the computation of Hamiltonian terms.
With CUDA-Q 0.6, developers can combine the capabilities of multi-QPU and multi-GPU approaches, allowing large-scale simulations to run in parallel. By merging the scalability of GPU-based circuit simulations with the parallelization enabled by multi-QPU mode, significant speedups are achievable. This integration empowers developers to experiment with various problem sizes and parallel endpoints to maximize GPU utilization effectively.
CUDA-Q leverages the Message Passing Interface (MPI) protocol for communication in parallel computing. With the integration of an MPI plugin interface, CUDA-Q can now seamlessly interface with various MPI implementations, including Open MPI and MPICH. This integration simplifies the adoption of CUDA-Q within existing MPI setups, enhancing interoperability and easing deployment in high-performance computing (HPC) centers or data centers.
Understanding Quantum Computing and Its Potential
Quantum computing, a field gaining significant traction in recent years, is a paradigm shift from classical computing. It leverages the principles of quantum mechanics to process information. The fundamental unit of quantum computing is the quantum bit or qubit. Qubits can exist in a superposition of states unlike classical bits, which can be either 0 or 1. However, this does not mean they represent both 0 and 1 simultaneously. Instead, the state of a qubit is a complex number representing a point in a two-dimensional space, often visualized as a sphere known as the Bloch sphere.
The power of quantum computing lies in the unique properties of quantum mechanics, such as interference and entanglement. Interference allows quantum algorithms to manipulate the probabilities of different outcomes, thereby guiding the computation towards the correct answer. This key property enables quantum computers to solve specific problems more efficiently than classical computers.
NVIDIA + Quantum, A formidable duality?
For those wanting to invest in frontier technologies such as Quantum Computing, there are pure play quantum computing companies such as IonQ, Rigetti and D-Wave, all with their specific quantum offerings, based around their technologies both hardware and software. The hardware part of the quantum stack can vary between different qubit technologies or even design pradiams, where D-Wave offers an annealing machine which differs from “gate” based quantum computers.
Then are the large technology companies, often listed companies, such as Google, Amazon, Honeywell, Microsoft and Intel who are all taking some bet on Quantum computing and associated technologies in some way. Of course, these companies have their bread and butter sources of revenue which enable them to invest in emerging, disruptive and frontier technologies. That offers some clear investment benefits. For example if Quantum goes through a “Quantum Winter“, some of the pure play quantum companies could fall by the way-side and go to zero as investors lose patience. However larger more diverse sources of income could allow major players to rise any waves of adoption of Quantum.
Could NVIDIA present something different? Its business is creating the hardware and software that enables AI. That could bleed over into Quantum AI. Quantum Machine Learning (QML) or any spin off that requires massive compute capacibilities is likely to have an edge on quantum technologies, for most part as large circuits are simulated before being run on actual hardware.
CUDA as generate product also is the go to for accelerated computing. Even rivals such as Intel have adopted CUDA as its so popular among developers. As developers embrace hybrid compuational workloads with differing types of compute, being a key advantage to NVIDIA is the massive installed base of users and developers. Could NVIDIA even purchase the specialist quantum hardware if it needs? It may not need to, just because its eco system puts it potentially as THE interface to whichever quantum device.
References
- Arute, F., Arya, K., Babbush, R., Bacon, D., Bardin, J. C., Barends, R., … & Chen, Z. (2019). Quantum supremacy using a programmable superconducting processor. Nature, 574(7779), 505-510.
- Sutor, R. S. (2019). Dancing with qubits: How quantum computing works and how it can change the world. No Starch Press.
- Shor, P. W. (1997). Polynomial-Time Algorithms for Prime Factorization and Discrete Logarithms on a Quantum Computer. SIAM Review, 41(2), 303–332.
- Preskill, J. (2018). Quantum Computing in the NISQ era and beyond. Quantum, 2, 79.
- Leibfried, D., Blatt, R., Monroe, C., & Wineland, D. (2003). Quantum dynamics of single trapped ions. Reviews of Modern Physics, 75(1), 281.
- Nielsen, M. A., & Chuang, I. L. (2010). Quantum computation and quantum information: 10th anniversary edition. Cambridge University Press.
- Svore, K. M., Hastings, M. B., & Freedman, M. H. (2018). Faster phase estimation. Quantum Information & Computation, 14(3-4), 306-328.
- Grover, L. K. (1996). A fast quantum mechanical algorithm for database search. Proceedings of the 28th Annual ACM Symposium on the Theory of Computing, 212–219.
- Harrow, A. W., Hassidim, A., & Lloyd, S. (2009). Quantum algorithm for linear systems of equations. Physical Review Letters, 103(15), 150502.
- Lloyd, S. (1996). Universal Quantum Simulators. Science, 273(5278), 1073–1078.
- Nvidia Corporation. (2021). Nvidia CUDA Toolkit Documentation. Nvidia Corporation.
- Monroe, C., & Kim, J. (2013). Scaling the Ion Trap Quantum Processor. Science, 339(6124), 1164–1169.
- Biamonte, J., Wittek, P., Pancotti, N., Rebentrost, P., Wiebe, N., & Lloyd, S. (2017). Quantum machine learning. Nature, 549(7671), 195-202.
- Feynman, R. P. (1982). Simulating physics with computers. International Journal of Theoretical Physics, 21(6–7), 467–488.
- Cook, S. (2012). CUDA Programming: A Developer’s Guide to Parallel Computing with GPUs. Morgan Kaufmann.