Using multiple colors of light to transmit data at the same time, new photonic chips break through the "bandwidth bottleneck" of high-performance computing

In the latest issue of "Nature Photonics", researchers from Columbia University School of Engineering demonstrated a new energy-saving chip that can transmit large amounts of data through fiber-optic cables connecting nodes. Instead of using multiple lasers to produce different wavelengths of light, the chip requires just one laser to produce hundreds of different wavelengths of light that can carry independent data streams simultaneously.

Photonic integrated links driven by Kerr frequency combs. 
Image credit: Light Wave Research Laboratory/Columbia University School of Engineering

In data centers and high-performance computers running artificial intelligence programs such as large language models, the amount of data they transfer between nodes is the source of the current "bandwidth bottleneck" that limits the performance and scalability of these systems.

Nodes in these systems can be more than a kilometer apart. These systems transmit data over fiber optic cables because metal wires dissipate electrical signals as heat when transmitting data at high speeds. Unfortunately, when signals are sent from one node to another, a lot of energy is wasted in the process of converting electrical data to optical data (and back again).

The newly developed millimeter-scale system uses wavelength division multiplexing and Kerr frequency comb devices to receive monochromatic light at the input end and generate many new colors of light at the output end. These devices are an ideal source of optical communications, where one can encode independent channels of information for each color of light and propagate them through a single optical fiber. This breakthrough could allow systems to transmit more data without using more energy.

The team designed a novel photonic circuit architecture that allows each channel to encode data independently with minimal interference to adjacent channels. In the experiment, the researchers successfully transmitted 32 different wavelengths of light at a speed of 16 gigabytes per second, with a total single fiber bandwidth of 512 gigabytes per second, and an error rate of less than 1 in 1 trillion bits of data. Bits – reaching incredibly high levels of speed and efficiency. The silicon chip that transmits the data measures just 4 millimeters by 1 millimeter, while the chip that receives the light signal and converts it into electrical signals measures just 3 millimeters by 1 millimeter, both smaller than a human fingernail.

The results demonstrate a feasible path to significantly reducing system energy consumption while increasing computing power by orders of magnitude, allowing AI applications to continue growing at an exponential rate with minimal environmental impact.