A photonic computing chip from Q.ANT (pronounced "
Q-ant") in Stuttgart, Germany that is used for AI training and inference. In partnership with the Institute for Microelectronics Stuttgart, the Native Processing Unit (NPU) is expected to have 30x more power efficiency than GPUs. The first photonic chip to achieve 8-bit precision, it is built with lithium niobate on top of silicon dioxide on top of silicon. A unique feature is that NPUs are fabricated in CMOS foundries, which provides a commercial outlet for older chip plants.
The NPU converts numbers into light frequencies, all of which travel together. It enables optical elements to perform multiplication using less energy than traditional digital processors.
Some Operations Are Much More Efficient
Fourier transforms are algorithms that are sometimes used in machine vision and model optimization. When used in digital computers, it takes a million transistors to run the algorithm, but when a photonic computer is used, only one device is required.
The Full Stack
Q.ANT provides the entire pipeline from design to finished wafers. Its Light Empowered Native Arithmetic (LENA) architecture unites the photonic and electronic worlds and works with traditional AI development tools such as TensorFlow and PyTorch. See
photonic processor.