Karlsruhe Institute of Technology (KIT) has begun operating a new computing test system that uses laser light to perform selected mathematical calculations instead of relying entirely on billions of electrically switched transistors. Developed by Stuttgart-based Q.ANT, the device is a photonic Native Processing Unit, or NPU, designed as a specialised accelerator that connects to conventional systems through PCIe and works alongside CPUs and GPUs rather than replacing them. Digital information is converted into properties of light and passed through a photonic integrated circuit, where interference between optical waves performs operations such as matrix multiplication. These calculations are fundamental to artificial intelligence, image recognition, scientific simulations and other data-intensive workloads. Germany’s Federal Ministry of Research, Technology and Space is providing around €4.3 million for the KIT test infrastructure, which will examine which parts of real-world scientific and AI applications can most efficiently be shifted from electronic processors to photonic hardware.
Independent testing at the Leibniz Supercomputing Centre has already produced striking results. Comparing the first two generations of Q.ANT processors, LRZ researchers found that the second-generation device could operate up to around 50 times faster in the best cases, while inference using a convolutional neural network for image recognition was accelerated by a factor of 25. For typical high-performance computing workloads, power consumption fell by roughly 50–84% between the two generations. The photonic PCIe cards themselves consume around 25 to 100 watts, while complete Q.ANT server systems require approximately 350–420 watts. Some current high-performance GPUs can draw up to around 1,000 watts, although this is not a direct like-for-like benchmark. Q.ANT makes much larger commercial claims – up to 90 times lower energy consumption per workload and up to 100 times greater data-centre capacity – but those figures come from the manufacturer and should be distinguished from independently measured results.
The potential matters because AI is rapidly becoming an electricity problem as well as a computing problem. The International Energy Agency estimates that data centres consumed roughly 415–460 TWh of electricity globally in 2024 and projects demand could reach around 945–1,000 TWh by 2030. The European Commission says data centres already account for about 2.5% of EU electricity consumption, while their installed capacity could rise from approximately 12 GW in 2025 to around 28 GW by 2030. Photonic processors are unlikely to replace silicon chips in the foreseeable future: they operate with analogue precision, results still have to be converted back into electrical signals, and many computing operations must remain on conventional hardware. But if future generations sustain the current gains, light could take over some of AI’s most energy-intensive mathematical workloads, allowing computing capacity to grow without electricity consumption rising at the same rate.
Sources: Karlsruhe Institute of Technology (KIT), Leibniz Supercomputing Centre (LRZ), Q.ANT, International Energy Agency (IEA), European Commission.
