Q.ANT Native Processing Unit (NPU)
Photonic co-processor on thin-film lithium niobate that performs AI matrix maths with light; second generation runs at LRZ and Jülich and is coming to IONOS cloud.
- Made by
- Q.ANT
- From
- Stuttgart, Baden-Württemberg
- Category
- Semiconductor equipment
- Introduced
- 2024
What it is
The Native Processing Unit is Q.ANT's photonic AI accelerator, packaged in a Native Processing Server that installs in an ordinary data-centre rack. It is a co-processor rather than a GPU replacement: it attaches through a standard PCIe interface and takes over specific mathematical workloads, chiefly the large matrix multiplications at the heart of neural networks.
How it works
Inside, photonic integrated circuits made from thin-film lithium niobate (TFLN) encode numbers in light and carry out operations in the optical domain. Lithium niobate is a crystal long used in telecom modulators because it switches light very fast; Q.ANT fabricates its chips on its own pilot line in Stuttgart, run together with the research institute IMS CHIPS. The company says the processors reach 16-bit floating-point accuracy, which matters because analogue optical computing has historically struggled with precision.
Why it stands out
Few photonic computing efforts anywhere have hardware running in production supercomputing centres. The second-generation NPU is installed at the Leibniz Supercomputing Centre (LRZ) and at the Jülich Supercomputing Centre. Evaluations at LRZ, reported by the company and trade press, found up to 50 times higher throughput than the first generation. Q.ANT's own benchmarks claim up to 30 times higher energy efficiency than conventional processors in equivalent matrix operations, measured at circuit level; treat that as a vendor figure until independent system-level data appears. Software support is moving too: in 2026 the compiler start-up Daisytuner compiled a PyTorch object detection model onto the hardware, and in June 2026 Q.ANT showed a diffusion model and NXAI's TiRex time-series model running on it.
Where you meet it
For now in research HPC and, from late 2026, through the IONOS public cloud, which announced in May 2026 that it would make Q.ANT servers available to commercial customers. Target workloads named by the company and LRZ include climate modelling, medical imaging and materials simulation.
Sources
Stories
Why Germany's deep-tech founders cluster around Munich, Aachen and Dresden
Three cities produce a striking share of Germany's hardware start-ups. The reasons are older and more specific than talent alone, and the failures of Lilium and Volocopter show the limits.
Twelve German deep-tech startups building real hardware
Photonic processors, e-fuel plants, wildfire satellites, graphene chips and heat batteries: twelve young German companies whose products you can weigh, install or launch.
Related products
AIXTRON MOCVD systems
Deposition systems from Herzogenrath that grow thin layers of compound semiconductors on wafers, used for LEDs, lasers and power electronics.
Infineon CoolMOS
High-voltage power MOSFETs released in 1998, the first commercial products based on the superjunction principle, which cut losses in power supplies.
TRUMPF EUV drive laser
A high-power CO2 laser system that fires 50,000 pulses a second at tin droplets, creating the plasma that emits the light used in EUV chipmaking.
Last reviewed 4 November 2026. Main source: www.globenewswire.com/news-release/2026/06/23/3316088/0/en/q. Suggest a correction.