Quantum computing faces the ghosts of its past, present and future

Quantum computing could be a disruptive technology. It’s founded on exotic-sounding physics and it bears the promise of solving certain classes of problems with unprecedented speed and efficiency. The problem, however, is that to this day, there has been too much promise and not enough delivery in the field, some say. Perhaps with the exception of D-Wave.

The company that pioneered quantum computing over 15 years ago has clients such as BASF, Deloitte, Mastercard and GlaxoSmithKline today. Alan Baratz went from running D-Wave’s R&D to becoming its CEO, taking the company public while launching products and pursuing new research directions.

In an exclusive interview, Baratz spoke to VentureBeat about quantum computing fundamentals and how this is related to the market’s current state, real-world clients and use cases, and what the future holds for this space.

Quantum computing hype and reality

Baratz has a diverse background that includes product management stints at Avia and Cisco, startup CEO stints and exits, as well as venture investment experience. What he considers closer to the work he is doing today with D-Wave, however, is being the first president of Javasoft at Sun Microsystems.

At Javasoft, Baratz was responsible for bringing the Java technology to market, building the developer ecosystem and growing revenue. As he noted, a lot of what he did there is similar to what D-Wave is doing now: creating a new industry and building a new ecosystem.

Of course, there are some fundamental differences. Java worked from Day 1, albeit not perfectly, and was built on existing infrastructure. From there, the technology grew, matured, and conquered the software development world. Last but not least, there is no fundamental technology divide in Java and, even though there may have been some hype and controversy around it at some point, it’s long been a proven technology.

Quantum computing, on the other hand, is a radically new concept which took years of R&D to develop and is not aimed at software developers. There is a fundamental technology divide in quantum computing, which Baratz explained is the source of both D-Wave’s dominant place in the market as well as the hype.

And yes, there’s lots of hype around quantum computing. According to McKinsey, the current state of quantum computing is between hype and revolution. According to the managing director of research at Bank of America, Haim Israel, quantum computing will be “bigger than fire.” According to quantum computing expert, Sankar Das Sarma, quantum startups are all the rage, but it’s unclear if they’ll be able to produce anything of use in the near future.

Baratz’s own position seems to be somewhere in between the above, drawing a line between quantum computing applications today and in the future, as well as between D-Wave and the competition.

“While everybody else in the quantum industry talks about government research grants as revenue and national labs and academic institutions as customers, we talk about companies like Mastercard, PayPal, GlaxoSmithKline, Johnson and Johnson, Volkswagen, BASF, Deloitte, SavantX and the port of L.A.,” said Baratz.

Quantum computing history and fundamentals

The dividing line between D-Wave and the competition that Baratz drew coincides with the line between the two different ways of building quantum computers: quantum annealing and gate models. As Baratz explained, when D-Wave embarked on the task to build a quantum computer over 15 years ago, it was thought that a gate model system could solve all problems. Quantum annealing, on the other hand, was known to only be able to address certain classes of problems.

There are four categories of problems that quantum computers can solve: optimization, linear algebra, factorization and differential equations. Baratz provided examples of applications for each: machine learning for linear algebra, cryptography for factorization and computational fluid dynamics and quantum chemistry for differential equations.

Optimization has a wide range of applications in physics, biology, engineering, economics and business. As Baratz noted, annealing quantum computers are very good at optimization problems. They can also solve linear algebra and factorization problems, but they cannot solve differential equation problems.

Back when D-Wave set out to build its quantum computer, the science and the engineering had not yet progressed to the point where it was believed that you could build a gate model system, Baratz explained. However, he added, it was widely accepted that you could build an annealing quantum computer. So D-Wave decided to go ahead and build an annealing system because that was something they believed they could do.

“Everybody else concluded that they might as well build a gate model system because they believed they [eventually] could and it could solve all problems, whereas annealing, it was known, could only solve a subset of the problems. So, everybody else jumped into gate. What happened was: a year ago, everybody got surprised, us included, because that’s the point in time at which it was proven that gate model systems can’t really deliver a speed-up on optimization problems”, Baratz noted.

Gate model systems are very good at differential equations problems. They can also attack linear algebra and factorization, but they cannot address optimization problems, Baratz said. In a nutshell, annealing can’t solve differential equations, while gate can’t solve optimization. As optimization has many potential applications, it turns out that’s pretty important.

D-Wave took what looked like a more conservative approach originally and was vindicated in retrospect. Baratz called this “a fluke of history that worked out really well for us.” By now, D-Wave has the first-mover advantage in annealing. This means they don’t just have expertise and technology others don’t, they also have a number of patents. All of that results in an effective moat for the company.

The problem with quantum computing

A year ago, D-Wave concluded that their annealing quantum computers had achieved commercial status. That means that they were capable of solving real business problems at commercial scale and “a lot, if not most, of the hard underlying technological problems had been solved,” as per Baratz. As the company had some bandwidth, they decided to initiate a gate model program that would allow them to eventually be able to address the full market for quantum.

Therefore, D-Wave also has firsthand experience of the issues gate model-based efforts are facing. The most severe one is dealing with errors. In conventional computing, bits are used for calculations and for storing information. The equivalent in quantum computing is qubits, and there is lots of talk about how many qubits each system can manage. The problem, however, is that more in this case does not necessarily mean better.

Qubits are much more sophisticated than bits, but there are many more ways that errors can be introduced, too. That typically happens by interacting with the environment, for example via electromagnetic interference. As Baratz noted, no system, quantum or otherwise, is error-free. In classical computers, we don’t usually think about errors because there are error-correction algorithms that take care of them. Quantum computers are not there yet.

Again, however, there are differences between annealing and gate model systems, according to Baratz. Gate model systems are very sensitive to errors, and that has to do with the way computation is performed. Doing a computation on a gate model system means applying instructions to qubits, similar to applying instructions to bits in classical computers. As soon as an error gets introduced, if it’s not corrected, the computation falls apart.

“Since these errors occur so frequently; without error correction, you can’t get through more than 20 or 30 instructions without the introduction of an error and the computation falling apart. But for many of the gate model algorithms, you need tens of thousands, hundreds of thousands or millions of gate instructions. So, you can’t do very much with a gate model system without error correction,” Baratz said.

Baratz sees error correction, not number of qubits or topology, as the key to enabling qubits that can have high fidelity through long computations and therefore making progress in the development of gate model systems. His estimate is that we are at least seven to 10 years away from reaching that point today. Annealing-based systems are much more stable, he said, although an increase in number of qubits and better qubit connection topologies would enable them to tackle more complex problems than what they can solve today.

Solving real-world problems

Baratz referred to fully optimizing FedEx routing from backbone to last mile as a problem that cannot be tackled today, as that would require tens of millions of variables. D-Wave is not there yet; however, a number of important real-world problems can already be solved. At the same time, progress is being made in terms of new computers with more qubits, better connectivity and lower error rates.

Baratz also referred to some of the problems that are being solved today, such as customer offer allocation for Mastercard, job scheduling for BASF and supply chain logistics with SavantX and the port of L.A. In that last use case, a 60% improvement in the performance of the cranes loading and offloading the containers and a 12% reduction in the time for vehicles to pick up goods was achieved.

Based on Baratz’s description, the philosophy of using gate model-based quantum systems sounds closer to programming classical computers. Using annealing-based quantum systems, however, is very different. There is no programming in the conventional sense involved. Tasks are modeled as optimization problems, which means that users need to declaratively state how their problems are defined, what are the parameters and their interdependence.

As Baratz noted, this is not something software engineers are expected to do, but rather something addressed by people like data scientists and data analysts. Optimization problems are often specified as what’s called a linear programming problem or a quadratic programming problem. This is the language that optimization engineers use, Baratz said, and D-Wave allows them to take that specification and feed it directly to hybrid solvers.

A hybrid solver utilizes both quantum and classical computers to solve problems. D-Wave has a hybrid solver in its offering, which recently got an upgrade. As Baratz described, the hybrid solver takes problem definitions as input and can determine which parts of the problem can be addressed by the quantum computer. It subsequently routes those parts of the problem to the quantum computer.

D-Wave’s offering, traction and roadmap

D-Wave offers a cloud service called Leap through which users can access its capabilities: quantum computers, hybrid solvers and software development tools. D-Wave also offers professional services to help clients with things like problem formulation or job submission, where expertise is not available in-house.

Given the current state of quantum computing, we wondered whether D-Wave’s clientele is made exclusively of the world’s largest companies. D-Wave is itself a publicly traded company listed on the New York Stock Exchange. As Baratz explained, by going public, D-Wave managed to raise cash and open up a variety of new funding sources.

In the call to discuss D-Wave’s recent Q3 results, which Baratz referred to as strong on all levels, the company announced that in the first three quarters of 2022, they had over 100 customers. Of those, 40 are government and education and 60 commercial, of which over 20 are Global 2000. D-Wave has around 40 commercial customers that are not Global 2000, Baratz said, such as a Canadian grocery chain called Save on Foods.

D-Wave’s core offering is also available via AWS Marketplace. In addition, D-Wave has a more targeted offering on AWS Marketplace: feature selection for machine learning. Feature selection is one of the most important elements of machine learning. When training a machine learning model, there will be a number of characteristics or classifiers that may be of interest to include. But including all of them will result in overfitting; i.e., generating a model that is not suited for the task at hand.

This is why a pre-processing step in machine learning is trying to identify a small set of representative characteristics and then building a model on that set. Finding a small set of strong classifiers from a big set of weak classifiers is a very hard optimization problem, and one in which D-Wave’s system does well. This is often used in fraud detection, Baratz said. Other parts of the machine learning process pipeline are not addressed by D-Wave at this point, because neither its quantum computer nor any of the gate model systems are yet capable of beating GPUs, according to Baratz.

Overall, Baratz concluded, the quantum ecosystem is defined by the annealing vs. gate models divide. Annealing is commercial today, while with gate models, things are still at a research and experimentation stage.

“We’re the only company in the world that does annealing to address optimization. Now we’re doing gate as well. So, we’ll be the only company in the world that can address the full market for quantum,” Baratz said.