AI Trends

Q.ANT Releases NPU Gen 2+ Photonic Chip, First Demonstration of Running Large Language Models

Semiconductor Industry PanoramaUpdated Nov 18 18:45

AI summary

German company Q.ANT unveiled its next-generation photonic processor, NPU Gen 2+, based on thin-film lithium niobate at the SC26 conference, and demonstrated on-site a hybrid system combining it with an AMD GPU running a large language model. The processor adds optical discrete Fourier transform capabilities, using optical operations to replace the computationally expensive attention layers in large models, aiming to alleviate computing bottlenecks. The company plans to launch a commercial system integrating both generations of products in the second half of 2027 and has open-sourced an SDK to support developer testing.

Why read this

AI score 60

Photonic processors enter large model workflows for the first time, demonstrating the feasibility and potential energy efficiency advantages of optical computing as a GPU co-processor.

Topics

Tags

Body · AI translation

Photon chips run large language models for the first time Source: Semiconductor Horizon Published: 2026-11-18 17:36 Optical computing begins to enter AI computing workflows.

Recently, German photonic computing company Q.ANT unveiled its next-generation photonic processor, NPU Gen 2+, at the SC26 Supercomputing Conference in Chicago, USA, and demonstrated a hybrid computing system composed of this processor and an AMD GPU running a large language model on-site. According to official information from Q.ANT, this marks the first time photonic computing has been involved in running large language models.

The NPU Gen 2+ is manufactured based on thin-film lithium niobate (TFLN) photonic integrated circuits and adds optical discrete Fourier transform (DFT) capabilities. In the demonstration, optical Fourier transforms replaced one layer of attention computation with high computational overhead in the large language model. The company plans to launch commercial systems integrating both NPU Gen 2 and Gen 2+ in the second half of 2027.

Fourier transforms enter the optical domain, yielding results in a single operation

The discrete Fourier transform is a fundamental operation that converts discrete signals from the time domain to the frequency domain, widely used in signal processing, image processing, and data analysis. Traditional digital processors achieve this transformation through multi-step numerical calculations; the larger the data scale, the higher the overhead.

The core of the NPU Gen 2+ is an optical device on the photonic chip dedicated to executing discrete Fourier transforms, with related computations completed in a single optical processing step. According to the official press release, this expands available operations beyond classical addition and multiplication, providing new algorithm design space for AI and high-performance computing applications.

Continuing the positioning of its predecessor, the NPU Gen 2+ serves as a co-processor alongside CPUs and GPUs, integrated into standard rack-mounted servers via PCIe interfaces. Q.ANT's approach is to assign the most compute-intensive specific workloads to hardware with the highest processing efficiency. This processor is an iteration of the NPU Gen 2, released in 2025, which primarily handled matrix multiplication and non-linear operations.

Optics replace the attention layer, hybrid system runs large model

English text is being prepared.

Google's research team developed the FNet model, which uses parameter-free Fourier transforms to replace the self-attention sub-layer in Transformer encoders. According to their paper data, FNet achieved 92% to 97% of the accuracy of BERT models of the same specification on the GLUE benchmark, improved training speed on GPUs by nearly 7 times, and doubled it on TPUs. However, FNet's Fourier transforms are still executed by digital processors; Q.ANT's difference lies in handing the transformation itself over to optical devices.

Michael Förtsch, CEO of Q.ANT, stated, "With the launch of Gen 2+, we have opened the door for software that leverages the natural capabilities of light, leading to more streamlined AI models with fewer parameters and less data. By running large language models in a hybrid system alongside AMD GPUs, we have demonstrated that this processor can function as a co-processor for GPUs."

Open-source SDK lowers barriers, pilot prepares for commercial use

Founded in 2018 and headquartered in Stuttgart, Germany, Q.ANT operates its own TFLN chip pilot line. According to official information, its systems have been deployed in multiple scientific computing centers and are offered for commercial deployment through cloud service provider IONOS.

In terms of development tools, Q.ANT open-sourced its Software Development Kit (SDK) in September 2026, providing C and Python interfaces, sample applications, documentation, and a CPU-based simulation environment. Developers can design and test photonic computing applications locally without physical optical hardware. For the NPU Gen 2+, the company is expanding this SDK to include discrete Fourier transform functionality, allowing developers to study the application of Fourier algorithms in AI and high-performance computing software.

Regarding hardware trials, Q.ANT launched a photonic computing pilot project with IONOS, offering bare-metal servers in German data centers to selected users for remote access to real photonic processors. The pilot currently provides NPU Gen 2 hardware. Utz Bacher, Vice President of Software at Q.ANT, said, "Photonic computing gives software developers the opportunity to rethink the mathematical logic behind applications. When complex mathematical operations can be completed directly through the physical properties of light, the questions developers can ask change."

Q.ANT also plans to combine multiple optical operations on a single photonic chip to support more complex computing processes. According to official statements, the scalability advantage of photonic computing ultimately depends on algorithm adaptation, system integration, computational precision, and end-to-end performance and energy consumption metrics.

The commercialization of photonic computing is moving from point demonstrations to a complete product path, with chips, software tools, cloud hardware trials, and planned commercial systems forming a connected line. The NPU Gen 2+ extends available optical operations from matrix multiplication to Fourier transforms, thereby bringing photonic processors into the workflow of large language models.

For purchasers of computing power in data centers, the actual value of such co-processors will need to be verified by practical test data after the delivery of commercial systems. Officials state that integrating photonic hardware into data center environments is expected to significantly reduce energy consumption and operating costs; whether this claim can be fulfilled is the main observation point for the subsequent development of this technology route.

This article was reposted from media reports or online platforms, reflecting the author's personal stance or views. Our republication is solely for sharing purposes and does not imply our endorsement or agreement. If the source is incorrectly marked or your legal rights are infringed, please contact customer service promptly. As a neutral platform service provider, we will correct, delete, or handle the matter according to law.

Comments: No user comments yet Latest News More Related Recommendations MLCC lead times extended to 10 months 2026-11-18 17:52

$150 million: South Korea's blank mask leader places its first overseas factory in Suzhou 2026-11-18 17:52

Can AI agents truly shift chip design left? 2026-11-18 17:46

Source

Semiconductor Industry Panorama