Industry-first Photonic AI Accelerator for energy-efficient High-Performance Computing and real-time AI Applications available in a 19″ rack-mountable Server
We unveil the second generation of our photonic processor to power the next wave of IA and HPC. Learn what is new for NPU Gen 2:
Join the path towards up to 30× higher energy efficiency and 50× faster computation, unprecedented compute density and bandwidth and radical reductions in operational costs for data centers.
The Q.ANT Native Processing Server (NPS) is the first commercial photonic processor for energy-efficient and accelerated AI and HPC workloads. By computing complex, mathematical functions with the natural properties of light, we realize significant performance and energy efficiency without digital detours while communicating seamlessly with existing computing infrastructure. Native Computing by Q.ANT promises:
Photonic Computing is reality: The Leibniz Supercomputing Centre (LRZ) and the Juelich Supercomputing Centre (JSC) – two of Europe’s leading HPC data centers – integrate Q.ANT’s NPS into their operational HPC environment. These first deployments mark a major step towards redefining how data centers approach performance, footprint and energy-efficiency.
Seize the exclusive opportunity to experience Q.ANT’s first commercial Photonic AI Accelerator, promising to set new standards in energy efficiency and computational speed. Test it, push it and see what is possible when AI runs natively on light. Get hands-on access to a completely new way of computing and redefine the possibilities of AI processing.
The Native Processing Server (NPS) is a 19″rack-mountable server with the Q.ANT photonic NPU PCIe card and is designed specifically for AI inference and advanced data processing. Its Plug & Play system design enables seamless integration into existing data centers and HPC environments, providing immediate access to photonic computing. The NPS is upgradable with additional NPU PCIe cards to increase processing power as workloads grow.
| System / Subsystem | Feature |
| System node | x86 processor architecture; 19” 4U commercially available rack system |
| Operating System | Linux Debian/Ubuntu with Long-Term Support |
| Network interface | 2x 10 Gbit ethernet, 1x 1 Gbit service interface |
| Software interface | C / C++ and Python API |
| API to NPU subsystem | Linux device driver |
| Native Processing Unit NPU |
|
| Power consumption of NPU | 150 W |
| Photonic integrated circuit (PIC) | Ultrafast photonic core based on z-cut Lithium Niobate on Insulator (LNoI) |
| Throughput of NPU | 8 GOPS |
| Operating temperature range | 15 to 35°C |
Each milestone brings us closer to unlocking the full potential of Photonic Computing: dramatically faster performance at a fraction of the energy.
Operation speed
Our photonic processors compute mathematical functions natively in light. This enables unprecedented throughput, projected to accelerate from 0.1 GOps in 2024 to 100,000 GOps by 2028 – a million-fold increase within five years.
Operation speed in GOps
Energy efficiency
Unlike transistors, photonic processors do not generate on-chip heat and use less components and parameters to solve complex tasks. This allows up to 30x lower energy consumption for AI workloads compared to conventional CMOS hardware, cutting both power demand and cooling requirements.
Energy consumption in fJ
We provide complimentary access to the Gartner® Hype Cycle™ for Data Center Infrastructure Technologies 2025 report. Learn how Photonic Computing can enable energy efficient data centers in a future of rising AI and HPC.
The Q.ANT Photonic Algorithms Library (Q.PAL) is the software interface to the NPS which enables users to operate directly at the multiplication level or to leverage optimized neural network operations such as fully connected layers or convolutional layers. Q.PAL offers a comprehensive collection of example applications that illustrate how AI applications can be enhanced. These examples can be implemented directly or as a foundation for creating custom use cases.
| Name | Description | Programming Language |
| Matrix Multiplication | Multiplication of a matrix and a vector | Python / C++ |
| Image Classification | Classification of an image (e.g. based on the ImageNet data set) | Python (Jupyter) |
| Semantic Segmentation | Segmentation of an image (e.g. based on a brain MRI scan data set) | Python (Jupyter) |
| Complex Line Fitting | Fitting of a high frequency line with a nonlinear network (e.g. based on simulated training data) | Python (Jupyter) |
Q.ANT’s photonic processing solution seamlessly integrates into the existing compute landscape. The Native Processing Unit at the heart of the NPS provides a PCIe interface housed in a standard 19” server, which makes the system plug-and-play. The NPU can be accessed via a software interface with C/C++ and Python APIs and will integrate into common AI frameworks such as PyTorch. Q.ANT supports customers in creating custom applications, providing the Q.ANT Photonic Algorithms Library and training resources.
Q.ANT’s Native Processing Server executes demanding AI workloads directly in light. While CMOS processors excel at linear, sequential processing, photonic processors are the natural hardware fit for large-scale nonlinear algorithms. Networks with nonlinear functions reduce the number of model parameters needed allowing higher accuracy per parameter and training budget. In photonic processors one single optical element performs one nonlinear operation while CMOS requires 100-1000 transistors and multiple cycles.
In this example, a network using learnable nonlinear functions on Q.ANT‘s NPS reconstructs complex image patterns more accuretaly than a linear network on a CPU while 2x less parameters and 3x less operations are needed.
# parameters: ~20k
# operations: ~670m
# parameters: ~10k
# operations: ~250m
Photonic Computing enables faster execution of matrix operations and non-linear functions directly in hardware. This allows for more efficient model architectures with fewer parameters. The result: higher throughput and lower power consumption for both, training large-scale models and real-time inference.
Many image processing tasks are inherently mathematical relying heavily on transforms like Fourier or convolution operations. With photonic processors, these operations can be performed optically, in parallel and at the speed of light. This dramatically increases frame rates and lowers energy usage.
Scientific simulations often depend on solving complex partial differential equations and large-scale matrix systems. Photonic hardware provides a powerful platform for these workloads by enabling high-bandwidth, computing that scales with problem complexity, helping simulate physical phenomena faster and with greater efficiency.
Q.ANT builds its photonic processors on a proprietary Thin-Film Lithium Niobate on Insulator (TFLNoI) platform. This material system enables the photonic integrated circuits (PICs) at the core of our NPUs. On a silicon wafer, a thin layer of lithium niobate is bonded to create optical waveguides, modulators, and other functional blocks — allowing high-speed, precise control of light on a single chip. We believe that TFLNoI is the key to the future of photonic computing.
PICs based on TFLNoI provide the physical foundation for scalable, efficient photonic computing delivering unique advantages:
SVP Native Computing
I look forward to discussing the potentials of Photonic Computing with you.
