Lithium-niobate photonic processor offers more efficient AI
Startup company Q.ANT GmbH (Stuttgart, Germany) has launched a commercial photonic processor based on lithium niobate electro-optic material in rack-mount and PCIe board-level formats.
The move to LiNbO3 allows many basic operations – such as matrix multiplication – to be done as a single analog operation and the company claims this brings at least 30x efficiency improvements and substantial performance boost compared with equivalent operations in digital CMOS. As such the LiNbO3 optical processors could one day replace silicon-based GPUs, such as those made by Nvidia
At the heart Q.ANT’s offering are thin-film lithium niobate (TFLN) on insulator die that contain configurable arrays of optical modulators. The die have multiple laser light inputs and provide multiple optical outputs and form what the company calls a Native Processing Unit (NPU).
To make access to the NPU seamless to users of conventional AI/ML servers Q.ANT has developed a firmware/software accompanying for electro-optic control and the complete stack is named LENA for Light Empowered Native Arithmetics. The product is compatible with existing computing ecosystem as it comes on the industry-standard PCI-Express. Or provider as a populated rack.
Big optical chip
The Q.ANT NPU executes complex, non-linear mathematics using light instead of electrons and is designed for compute-intensive applications such as AI Inference, machine learning, and physics simulation.
The NPU is designed to the full reticle size of 20mm by 30mm. It typically has 8 inputs and 8 outputs and in between an array of optical modulator, such as Mach-Zehnder interferometers Complexities can range from a 4 by 8 grid up to 800 modulators, Michael Förtsch, CEO of Q.ANT, told eeNews Europe.
The fact that optical paths are controlled by voltages – electrical fields – and not current means there is considerably less power consumption and much higher frequencies of operation can be achieved than within silicon photonics, said Förtsch.
The power (consumption) of GPT4
“With our photonic chip technology now available on the standard PCIe interface, we’re bringing the incredible power of photonics directly into real-world applications. For us, this is not just a processor—it’s a statement of intent: Sustainability and performance can go hand in hand,” said Förtsch in a statement. “For the first time, developers can create AI applications and explore the capabilities of photonic computing, particularly for complex, nonlinear calculations. For example, experts calculated that one GPT-4 query today uses 10 times more electricity than a regular internet search request. Our photonic computing chips offer the potential to reduce the energy consumption for that query by a factor of 30.”
Q.ANT has been developing TFLN since the company’s foundation in 2018. TFLN is not readily available in commercial fabs or from foundries and so Q.ANT has built its own pilot line in Stuttgart where it runs 4-inch diameter silicon wafers to support the TFNL. Förtsch told eeNews Europe the company is producing about 1,000 wafers per year with a plan to ramp production on 6-inch wafers in 2025. Ultimately, Förtsch said it would be good to find a source of optical grade TFNL-on- insulator on 200mm-diameter wafers.
The devices laid down in the TFLN are of micrometer scale but geometrical scaling is not so important for optical processing which can implement highly efficient parallel operations. Q.ANT gave the example of a Fourier transform that requires millions of transistors in traditional computing but can be accomplished within a single optical element. In addition, it is possible to run several operations in parallel using wavelength division multiplexing.
Yole’s Mounier says
There are numerous places in established AI/ML routines, both within neural networks operations and for transitions between, where optical processing can reduce complexity. Test runs on Q.ANT’s NPU demonstration system with MNIST datasets show accuracy comparable to linear networks while using 43 percent fewer parameters and reducing operations by 46 percent, establishing it as a more efficient choice for AI inference.
However, while TFLN may be good for replacing silicon GPUs established AI/ML ca
Eric Mounier, photonics & sensing analyst at Yole Group, said: Q.ANT’s novel approach to photonic processing is a ground-breaking step towards addressing the escalating energy demands of the AI era.” He continued: “This new processor generation finally gives access to superior mathematical operations, which have been too energy-demanding on traditional GPUs. The first impact is expected in AI inference and training performance, paving the way for high-efficiency, sustainable AI computing.”
The Q.ANT NPU is set to be available in February 2025 as a turnkey Native Processing Server (NPS) compatible with x86 systems and ready to plug into high performance computing (HPC) and data centers. The company also provides a Q.ANT software toolkit to ease integration with AI software stacks.
Related links and articles:
News articles:
Single carrier 400G/Lambda lithium niobate modulator
Nvidia’s Blackwell AI GPUs overheating, says report
If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :
