April 21, 2026 -- LRZ experimented with photonic processors from Q.ANT in everyday operations. After working with the novel technology, LRZ staffer see the technology as having strong development potential.

The display might show colorful waves moving from left to right across the screen, but cutting‑edge technology is at work behind this colourful decoration. Specifically, Q.ANT’s photonic‑analog accelerators, which the start-up calls Native Processing Units (NPUs), use laser light, which photodetectors convert into digital electrical impulses when results need to be processed further by other components of a computer system. Recently, the Leibniz Supercomputing Centre (LRZ) has been experimenting with Q.ANT’s technology, and according to the center, the results are promising. “We evaluated the first two generations of these Q.ANT processors and compared their performance and potential uses in data centers, methods of artificial intelligence (AI), and high‑performance computing (HPC),” says Dr. Josef Weidendorfer, computer scientist of the Future Computing research team at LRZ. “In principle, they represent a highly promising technology that can significantly increase energy efficiency in data centers, especially for training and inference of AI applications.”

Photonic accelerators promise solutions for both the rapidly increasing energy demand of AI clusters and for performance gains in computations such as matrix‑vector multiplications or nonlinear equations, which are widely used in HPC. Q.ANT discovered the computing power of light during its search for new computing technologies. Founded in 2018, the Stuttgart‑based startup now has production capacity for stringing together small series of its innovative co‑processors – a prerequisite for allowing potential users to test them. To this end, the Federal Ministry for Research, Technology and Space (BMFTR) financed a research procurement so that the LRZ could also explore the potential of photonic computing for science and industry.

Test Tasks from HPC and AI

Q.ANT supplies its analog‑photonic chips as plug‑in cards (PCIe) for computer mainboards and currently also in its own servers. “The co‑processors cannot perform all operations using light,” explains LRZ researcher Dr. Ehab Saleh. “Therefore, a continuous data exchange between photonic and classical processing units is required. This is why so-called complementary metal-oxide semiconductor or CMOS technology is also included on the cards.” The package additionally contains a function and software library: “Users don’t need to understand exactly how the accelerator works,” Saleh adds. “They can easily integrate library functions into their C/C++ or Python applications. These functions abstract the low-level communication details between the CPU and the accelerator, such as device initialization and release, data movement, state management, and access to auxiliary services.”

While current graphics processing units (GPUs) draw up to 1,000 watts of power, Q.ANT’s systems – consisting of server and card – require only 350 to 420 watts, the cards only 25 and 100 watts. For the tests at LRZ, three servers were procured to evaluate not only performance but also parallelization of photonic nodes. While the first-generation servers contained a single card each, their successors were equipped with three PCIe — an approach to increase performance. In addition to various computational workloads, the processors of both generations were tested with typical AI tasks such as pattern and image recognition as well as the training and inference of smaller neural networks (MNIST, ReSet). LRZ staff measured metrics such as execution time, performance, energy consumption, error rates, prediction reliability, and training loss.

Higher Performance and more Efficiency

Ultimately, LRZ researchers found the results are encouraging: while the optical element of the first generation performed around 100 million operations per second (Mop/s), the next generation achieved 500 Mop/s at each channel. Q.ANT also integrated eight lasers and optical channels into these newer cards. For the total that consequently amounts to four trillion operations per second or 4 Gop/s. “In the best case, the second‑generation chips operated 50 times faster,” Weidendorfer said. A convolutional neural network accelerated image recognition by a factor of 25 thanks to this improvement.

The evaluation team indicated that future development stages for the optical elements would likely focus on frequency and workload management. The current-generation chip uses a two-gigahertz frequency, which could rise to 50 or 100 gigahertz if a given system’s electronics can support the increase. They also suspect that future cards could accommodate thousands or even tens of thousands of optical processing units: In addition, using different wavelengths of light could further increase computing power, and users could expand performance by adding more cards. Further, performance gains do not come at the expense of energy efficiency. For typical HPC workloads, power consumption dropped by about 50 to 84 percent from generation 1 to generation 2, depending on computational complexity; in more demanding tasks. “Energy efficiency is highest when computations remain on the optical unit as long as possible,” Weidendorfer notes. “Otherwise, components on the chip should be placed as close together as possible to minimize energy used for data transfer.”

Like GPUs, analog‑photonic co‑processors do not compute with exact precision: they convert digital and analog data with 16‑bit precision, but the noise generated by photodetectors must be taken into account. “AI applications operate well with lower precision, which makes analog‑photonic computing particularly suitable for them,” Saleh observed. At LRZ, researchers believe that Q.ANT chips could be ready for deployment in AI clusters within one or two further development cycles – boosting both energy efficiency and performance. Weidendorfer and Saleh also indicated that the processor can also emulate higher precision and accelerate supercomputing. For more precise results, complex tasks – similar to hand calculations – must be broken into multiple operations, which requires further and more sophisticated software tools. However, to achieve more precise results, complex tasks—such as mental arithmetic—would need to be broken down into multiple calculations, which requires further and more complex enhancements to the included software tools. A key step in the further development of photonic computing would be the emergence of a more diverse programming environment for this technology, featuring extensive libraries, languages, and additional software: “For evaluating next-generation systems,” the researchers add, “it would be interesting to work with real application programs and optimize them for photonic processors. This would allow us to identify where the function library needs to be adapted and expanded.”