TL;DR
At SC26 in Chicago, Q.ANT demonstrated its NPU Gen 2+ photonic processor working alongside an AMD GPU to run a language model. The company says optical discrete Fourier transforms replaced a computationally intensive attention layer in the demonstration; performance, energy use and model-quality results were not provided in the source report.
Recommended by digitechbytes.com · AI-assisted guides
Shopping for technology news & gadgets? Start with the guides we keep up to date:
Q.ANT demonstrated its NPU Gen 2+ photonic processor working alongside an AMD GPU to run a language model at the SC26 conference in Chicago. The company said the system used optical discrete Fourier transforms to replace a computationally intensive attention layer, showing one way photonic hardware could act as a co-processor for AI workloads; the demonstration’s performance and energy results were not included in the report.
The NPU Gen 2+ adds the ability to perform discrete Fourier transforms (DFTs) directly in light, expanding the mathematical operations supported by Q.ANT’s photonic processors. Q.ANT said an optical component on the chip performs the transform in a single operation, rather than through a sequence of electronic calculations. DFTs are used in areas including signal and image processing, simulations and data analysis.
At SC26, the company paired the processor with an AMD GPU in a hybrid system running a large language model. Q.ANT described the demonstration as the first time it had run an LLM with optics and said the photonic processor handled calculations that replaced an attention layer. The report does not identify the model, describe the exact workload or provide benchmark figures, so the demonstration establishes a system configuration rather than a quantified performance advantage.
Q.ANT is extending its software development kit to support the Gen 2+ functions, including Fourier-based algorithms. The company launched the SDK in September 2026 with C and Python APIs, examples, documentation and a CPU-based simulation environment, according to the source report. Q.ANT also said a photonic pilot with IONOS provides access to physical hardware. It plans commercial systems combining Gen 2 and Gen 2+ processors in the second half of 2027.
Photonic Chips as AI Co-processors
The demonstration matters because it presents photonic computing as a possible specialized partner to GPUs, rather than a replacement for conventional processors. If a photonic chip can efficiently handle particular mathematical operations, developers might distribute workloads across different hardware types. Attention is one of the computationally demanding parts of language-model systems, so exploring alternative ways to implement it could have implications for how AI workloads are designed.
Those implications remain prospective. Q.ANT says its optical DFT capability may support leaner models and lower energy use, but the source report offers no measurements of speed, power consumption, cost or model accuracy. The demonstration therefore signals a technical direction, not evidence that photonic systems already provide a broad or commercially proven advantage for language models.
From Gen 2 to Optical DFTs
The NPU Gen 2+ builds on Q.ANT’s NPU Gen 2, launched in 2025 with a focus on matrix multiplication and nonlinear operations. The new generation adds DFTs, extending the set of operations the company says its processors can perform directly using light. Q.ANT’s stated approach is to use photonic processors as co-processors alongside CPUs and GPUs, assigning selected calculations to the hardware suited to them.
The software tools are part of that development path. Q.ANT’s SDK is intended to let developers create and test applications in a CPU simulation environment before accessing optical hardware. Its planned commercial systems will combine Gen 2 and Gen 2+ processors, but the source report does not specify system configurations, availability by market or pricing.
“With Gen 2+, we’re opening the door to software that uses what light does naturally.”
— Michael Förtsch, Q.ANT chief executive
Performance Data Still Missing
The report does not give benchmark results for the language-model demonstration, including runtime, throughput, energy consumption or comparison with a GPU-only system. It also does not name the model, describe the attention layer in detail, or state whether the model’s output quality was evaluated. Without those details, the scale of any practical improvement cannot be assessed.
Q.ANT’s statements about fewer model parameters, reduced data requirements and lower operating costs are company claims about potential outcomes, not reported measurements from the demonstration. The source does not say whether the system was tested by independent researchers or customers, or when developers will receive access to the Gen 2+ hardware and updated SDK.
SDK Work and 2027 Systems
Q.ANT says it is extending its SDK for the Gen 2+ DFT functions so developers can experiment with Fourier-based algorithms and their use in AI and high-performance computing. The next useful evidence would include technical documentation and reproducible comparisons showing how the optical operation performs within complete applications.
The company plans to offer commercial systems combining Gen 2 and Gen 2+ processors in the second half of 2027. The source report does not provide a specific release date or pricing. Further demonstrations and developer access may clarify which workloads can use the hardware and how the hybrid systems compare with conventional alternatives.
Key Questions
What did Q.ANT demonstrate at SC26?
Q.ANT demonstrated a language-model workload running on a hybrid system that paired its NPU Gen 2+ photonic processor with an AMD GPU. The company said the photonic processor used optical DFTs to replace a computationally intensive attention layer.
What is new in the NPU Gen 2+?
The processor adds discrete Fourier transforms performed directly in light. Q.ANT says this expands the operations developers can run on its photonic hardware beyond the functions supported by the earlier NPU Gen 2.
Did the demonstration prove the system is faster or more energy-efficient?
No such measurements are included in the source report. Q.ANT described possible performance and energy benefits, but benchmark figures and comparisons were not provided.
When does Q.ANT expect commercial systems?
Q.ANT plans to offer systems combining Gen 2 and Gen 2+ processors in the second half of 2027. The report does not give a more specific release date or pricing.
Source: thequantuminsider.com
