Calculating with Light Q.ANT Demonstrates Generative AI and xLSTM on Photonic Hardware

Related Vendors

Taiwan Machine Tool & Accessory Builders' Association (TMBA)

SIMTOS (Seoul International Machine Tool Show)

Q.ANT is making the transition from the lab to production-ready applications. The company is demonstrating how its second-generation photonic Native Processing Unit (NPU) handles complex AI workloads such as diffusion models and recurrent networks (xLSTM).

Photonics computing has reached a level of maturity that goes beyond basic algorithmic demonstrations. At this year’s ISC in Hamburg, Q.ANT is demonstrating that its proprietary NPU architecture supports the full range of modern AI applications: from generative image synthesis to sequential time-series forecasting.

Linear Algebra Using Light Instead of Transistors

At the heart of the demonstration is the processing of highly complex AI workloads directly on photonic hardware. Unlike traditional transistors, the NPU performs mathematical operations (linear algebra) using light signals. Dr. Michael Förtsch, founder and CEO of Q.ANT, sees this as the solution to the industry’s most pressing problem: “Photonics architecture changes the energy equation for the AI ecosystem. Energy is the bottleneck for the future of AI. Computing with light instead of transistors reduces energy consumption right at the source.”

At the chip level, these HPC workloads aim to achieve up to 30 times greater energy efficiency than traditional processors for equivalent matrix operations.

Diffusion Models And Generative Synthesis

To demonstrate its suitability for generative AI, Q.ANT is showcasing a diffusion model for image-to-image synthesis at the trade show. These workloads are considered particularly computationally intensive, as they involve iterative, massively parallelized matrix operations. The images are generated through repeated forward passes of a deep neural network.

Prof. Dr. Björn Ommer, head of the Computer Vision & Learning Group at LMU Munich and developer of diffusion models, underscores the significance of this breakthrough: “Diffusion models rely on very extensive computational operations to gradually generate a coherent output. If photonic hardware could execute such workloads efficiently and reliably, this would be an exciting indication that alternative computing architectures will play an important role in the future of generative AI.”

xLSTM And Time Series Forecasting

As further evidence of the architecture’s versatility, Q.ANT also presents the TiRex time-series forecasting model. This model is based on the xLSTM (Extended Long Short-Term Memory) architecture, which is commercialized by the Austrian frontier AI lab NXAI.

Unlike transformer-based models, xLSTM is a recurrent neural network (RNN) that recognizes patterns in sequential data and enables predictions over long time horizons. Lukas Fischer, Head of Applied Research at NXAI, emphasizes: “xLSTM architecture on photonic systems could redefine what energy-efficient AI actually means.” The TiRex model is specifically optimized for enterprise applications such as financial market analysis, supply chain optimization, and traffic flow simulations.

The Growing Ecosystem

A key factor for industrial use is integration into existing software workflows. Q.ANT highlights significant progress in this area:

• Compiler Integration: In collaboration with our partner Daisytuner, we were able to compile an object recognition model directly from the standard PyTorch framework for the photonic hardware as early as the beginning of the year.
• Cloud Infrastructure: In May, a strategic partnership was formed with the cloud provider Ionos to make photonic computing commercially available.
• HPC Deployment: Leading centers such as the Leibniz Supercomputing Center (LRZ) in Munich and the Jülich Supercomputing Center are already operating the hardware in production.

With these demonstrations, Q.ANT is proving that photonic computing has the potential to become an integral part of the next generation of computing—supported by an increasingly stable ecosystem of hardware, compilers, and cloud partners. (heh)

Partnership in Photon Processors

Ionos Relies on Photonic Chips from Q.ANT