British quantum computing company Orca is working with Nvidia to speed up and improve machine learning processes. The new hybrid system sends some of the processing to quantum processors, which improves output quality and reduces training time, the company says.
The partnership comes amid concerns about computational bottlenecks for training new large-scale language and underlying AI models. As models grow larger and their outputs more complex, so does the computational demand.
In a hybrid system, a combination of hybrid classical and quantum algorithms can be split between Nvidia GPUs traditionally used in machine learning and Orca’s new Quantum Processing Unit (QPU). William Clements, Head of Machine Learning at Orca Computing, said: tech monitor: “It’s best to think of the QPU as assisting the GPU in the training process.”
Orca has focused on quantum machine learning image generation and analysis technology. This is the type of his AI model that powers tools like Stable Diffusion and Midjourney. In addition to creating images from text prompts, this model can also create synthetic images for medical use and look for changes in environmental factors across large image libraries.
“What we do in generative modeling is how [QPU performance] We look at scale and how it affects the performance of the generation algorithm,” says Clements. “We are very excited about this hybrid approach, which will allow us to scale to very large systems. .”
He said it has demonstrated the ability to produce very high quality images using a combination of photonic QPUs and eight large GPUs. They found it to be faster and an improvement over GPU clusters working alone. “It’s there to help classical systems learn to approximate distributions. In the case of images, it’s the distribution of pixels. It’s about providing a rich distribution,” he explained Clements.
This effectively allows the QPU to calculate the best placement for each individual pixel generated by the GPU, tell the GPU where to place those pixels, and then the GPU cares nothing about producing high-quality images. It is meant to allow the work of generating an image to be performed. position and layout.
This technique can prove very valuable in other areas where a large number of small items need to be arranged in the most efficient way, such as placing transistors on a chip.
Content from partners
Quantum machine learning could solve computational bottlenecks
“Machine learning has come a long way in the last few years,” says Clements. “But there are still some issues. For example, it is very hungry in terms of data and computational resources. , is not necessarily so good for scientific datasets, molecules, etc.
“What quantum computing comes into play is that quantum computers can solve certain kinds of mathematical problems that classical computers cannot.”
Nvidia is focused on building the bridge between classical and quantum through its open-source CUDA Quantum toolkit and quantum simulation tools running on high-performance GPUs.
Companies like Orca are still working to develop a pure quantum computer with hardware powerful and reliable enough to power business, but it could be decades away. There is a nature. On the other hand, classical computers need to make up for that headroom, and hybrid models represent a viable and faster path to realizing some commercial quantum advantage.
Orca leverages this hybrid solution with generative adversarial network (GAN) type machine learning. In classical GANs, two neural networks are trained to compete against each other. One is a generator that transforms random numbers from a probability distribution called the “latent space” into data such as images. The other is an identifier that determines if the data is real or fake.
In a hybrid system, the “latent space” is provided by a photonic quantum computer, taking it away from the GPU, which is responsible for image generation and evaluation. “The quality of data produced by a generator is highly dependent on the type of latent space used,” he says Orca. “Using a quantum latent space improves the performance of his GANs on some datasets, such as those used in quantum chemistry.”
Orca’s PT series photonic quantum processors acted as the latent space during the experiment, and eight Nvidia A100 GPUs trained classical neural networks acting as generators and discriminators. All machines were on site.
“Hybrid quantum-classical computing has the potential to reimagine how industry leaders work together to solve some of the world’s toughest challenges,” said Timothy Costa, Director of High Performance Computing and Quantum at Nvidia. It’s hidden,” he said. “CUDA Quantum enables Orca’s PT-series photonic quantum processors to seamlessly develop and integrate into hybrid he workflows, opening a new era of quantum computing and making quantum computing more accessible to researchers than ever before.” It becomes easier to use.”
