Photonic computing First commercial photonic processor for energy-efficient high-performance computing

Related Vendors

Q.ANT GmbH

Wuxi InfiMotion Technology Co., Ltd.

With the new Native Processing Unit (NPU) from Q.ANT, electronics developers benefit from up to 30 times higher energy efficiency. Whether AI inference, machine learning, or physical simulations—light instead of electrons enables more sustainable and powerful applications.

At a time when the energy demand of data centers and AI systems is increasing exponentially, the development of sustainable technologies is coming into focus for the electronics and software industry. Traditional CMOS technologies are increasingly reaching their limits, especially when it comes to energy-intensive applications such as machine learning or complex physical simulations. Photonics is a key technology here and promises a fundamental change in computing technology.

With its first commercial photonic processor, Q.ANT not only delivers better energy efficiency but also significantly boosts performance with the first Native Processing Unit (NPU). This is especially relevant for data centers. The technology is designed for demanding applications such as AI inference, machine learning, or physics simulations.

The photonic architecture of the NPU is based on the company's proprietary LENA platform (Light Empowered Native Arithmetics) and utilizes the material thin-film lithium niobate (TFLN). Unlike conventional processors, the NPU performs calculations with light. This technology enables a drastic reduction in energy consumption—up to 30 times more energy-efficient than CMOS technologies—while also offering a significant performance enhancement.

Photonic computing reduces energy consumption

"With our photonic chip technology, now available through the standard PCIe interface, we are bringing the incredible power of photonics directly into real-world applications. We are making a clear statement: performance and sustainability can go hand in hand," says Dr. Michael Förtsch, CEO of Q.ANT.

"For the first time, developers can create AI applications and explore the possibilities of photonic computing, especially for complex, nonlinear calculations. Experts have calculated, for example, that a GPT-4 query today consumes ten times more power than a regular internet search query. Our photonic chips offer the potential to reduce the energy consumption for this query by 30 times."

Example: A Fourier transform, which requires millions of transistors, is performed with a single optical element at Q.ANT. This not only saves space on the chip but also significantly reduces power consumption. For electronics developers, this means more flexibility in system architecture without compromising computing power.

Q.ANT's novel approach to photonic processing is a groundbreaking step towards meeting the growing energy demands of the AI era.

Dr. Eric Mounier, Chief Analyst, Photonics & Sensing at the research firm Yole Group

Efficiency and sustainability in focus

Especially in the era of increasingly complex AI applications, the Q.ANT NPU offers critical advantages.

• The Q.ANT NPU can reduce the computational requirements for machine learning, computer vision, or for training and inference of large language models (LLM).

• Test runs with the NPU demo system in the cloud using MNIST datasets showed that Q.ANT's native computing approach achieves accuracy comparable to linear networks with less energy consumption.

• Simulations of Kolmogorov-Arnold Networks (KAN) also demonstrated that 43 percent fewer parameters are needed, and the number of operations can be reduced by 46 percent, establishing it as a more efficient choice for AI inference.

Further tests and simulations for image recognition show that the Q.ANT NPU can train significantly faster and achieve accurate recognition with only 0.1 million parameters and 0.2 million operations. A conventional approach struggles to achieve acceptable results even with 5.1 million parameters and 10 million operations.

It also enables faster solutions for partial differential equations in physics simulations, simplifies time series analysis, and improves efficiency in solving graph theory problems. Unlike standard CMOS technology, the Q.ANT NPU processes data using light, allowing for more energy-efficient mathematical operations. While a conventional CMOS multiplier requires 1,200 transistors to perform a simple 8-bit multiplication, the Q.ANT NPU achieves this with a single optical element. (heh)

Photonic components

Ansys, TSMC, and Microsoft multiply simulation by tenfold