Analysis Quantum computing has remained a decade away for over a decade now, but according to industry experts it may hold the secret to curbing AI's insatiable appetite.

With each passing month, larger, more parameter-dense models appear and the scale of AI deployments expand in tandem. This year alone hyperscalers like Meta, plan to deploy hundreds of thousands of accelerators. Even still OpenAI founder Sam Altman is convinced we’ll need exponentially more compute if we’re to further to develop of AI.

Hence it should come as no surprise that with its latest hire, OpenAI would be hedging its bets on quantum computing on the off chance it can. Last week, the AI juggernaut added Ben Bartlett, a former quantum systems architect at PsiQuantum to its ranks.

We reached out to Open AI to find out more about what Bartlett will be doing at the AI trendsetter, and haven't heard back. However his bio offers some hints as much of his research has focused on the intersection between quantum physics, machine learning, and nanophotonics, and "basically consists of me designing little race tracks for photons that trick them into doing useful computations"

So what exactly could OpenAI want with a quantum physicist? Well, there are a couple possibilities ranging from using quantum optimization to streamline training datasets or using quantum processing units (QPUs) to offload complex graph databases, to using optics to scale beyond the limits of modern semiconductor packaging.

Neural networks are just another optimization problem

Quantum computing has the potential to drastically improve the efficiency of training large AI models, allowing them to derive more accurate answers from models with fewer parameters, D-Wave's Murray Thom tells The Register.

With GPT-4 rumored to be in excess of a trillion parameters, it's not hard to see why this might be attractive. Without resorting to quantization and other compression strategies, AI models need about 1GB of memory for every billion parameters when running at FP8 or Int8 precision and at higher precisions, substantially more than that.

Trillion parameter models are nearing the limits of what a single AI server can efficiently accommodate. Multiple servers can be strung together to support larger models, but leaving the box imparts a performance penalty.

And that's today. And if Altman is right these models are only going to get bigger and more prevalent. As such, any technology that could let OpenAI increase the capability of its models without also increasing parameter count meaningfully could give it a leg up.

"As you're training a model, the number of parameters that go into the model really drives the cost and the complexity of training the model," Trevor Lanting, D-Wave VP of software and algorithms tells The Register.

To get around this, he explains, developers will often sub-select features they think are going to be the most important for training that particular model, which in turn reduces the number of parameters required.

But rather than trying to do this using conventional systems, D-Wave makes the case that quantum optimization algorithms may be more effective at determining which features to leave in or out.

If you're not familiar, optimization problems, like those commonly seen in path finding or logistics have proven to be one of the most promising applications of quantum computing thus far.

"What our quantum computers are really good at is optimizing things where things are either happening or not happening: like someone being assigned a particular schedule or being assigned a particular delivery," Thom said. "If those decisions were independent, that would be fine, and it would be easy for a classical computer to do, but they actually affect the other resources in the pool and there's sort of a network effect."

In other words, the real world is messy. There might be multiple vehicles on the road, road closures, weather events and so on and so forth. Compared to classical computers, the unique attributes inherent to quantum computers allow them to explore these factors simultaneously to identify the best route.

This, "is completely analogous to a neural network where the neurons are either firing or not firing, and they and they have synaptic connections to the other neurons, which either excite or inhibit the other neurons from firing," Thom explains.

And this means that quantum algorithms can be used to optimize AI training datasets for specific requirements, which when trained, results in a leaner, more accurate model, Lanting claimed.

Quantum sampling and offloading

Longer term, D-Wave and others are looking for ways to implement QPUs deeper into the training process.

One of these use cases involves applying quantum computing to sampling. Sampling refers to how AI models, like LLMs, determine what the next word, or more specifically token, should be based on a distribution of probabilities. This is why it's often joked that LLMs are just autocomplete on steroids.

"The hardware is very good at producing samples, and you can tune the distribution, so you can tune the weighting of those samples. And what we are exploring is: is this a good way to actually insert annealing quantum computing hard and more directly into the training workload," Lanting explained.

French quantum computing startup Pasqal has also been toying with applying quantum computing to offload graph structured data sets commonly found in neural networks.

"In machine learning there is no real simple way of representing the data classically, because the graph is a complex object," Pasqal Co-CEO Loïc Henriet explained in an interview with The Register. "You can embed graph structured data into quantum dynamics relatively naturally, which gives rise to some new ways of treating those pieces of data."

However, before this can be achieved, quantum systems are going to have to get a lot bigger and a lot faster, Henriet explained.

"Large datasets are not practical for the moment," he said. "That's why we are pushing the number of qubits; the repetition rate. Because with more qubits you can embed more data."

Just how long we'll have to wait before quantum graph neural networks become viable its hard to say. Pasqal already has a 10,000 qubit system in the works. Unfortunately, research suggests that it'll take more than 10,000 error-correcting qubits, or about a million physical qubits just to compete with modern GPUs.

A silicon photonics play?

Exotic quantum AI use cases aside, there are other technologies OpenAI could be pursuing for which Bartlett just so happens to be an expert.

Most notably, Bartlett's former employer PsiQuantum has been developing systems based on silicon photonics. This suggests his hire could be related to OpenAI's reported work on a custom AI accelerator.

Several silicon photonics startups, including Ayar Labs, Lightmatter, and Celestial AI have pushed the technology as a means to overcome bandwidth limits, which has become a limiting factor scaling machine learning performance.

The idea here is you can push a lot more data over a much longer distance with light than you can with a purely electrical signal. In many of these designs, the light is actually carried by wave guides etched into the silicon, which sounds an awful lot like "designing little race tracks for photons."

Lightmatter believes this technology will allow multiple accelerators to function as one without incurring a bandwidth penalty for data leaving the chip. Meanwhile Celestial sees an opportunity to vastly increase the amount of high-bandwidth memory available to GPUs by eliminating the need to co-package the modules directly adjacent to the accelerator die. Both of these capabilities would be attractive to a company working with AI systems at a massive scale.

Whether OpenAI will ultimately pursue quantum AI or silicon photonics remains to be seen, but for a company whose founder is no stranger to making long-shot investments, it wouldn't be the strangest thing Altman has backed. ®

https://www.theregister.com//2024/03/13/quantum_openai_hire/