English

The research team from the School of Engineering at Columbia University in the United States has broken through the "bandwidth bottleneck" of high-performance computing in new photonic chips

1049
2023-10-31 11:28:37
See translation

When running various artificial intelligence programs such as large language models, although data centers and high-performance computers are not limited by the computing power of their individual nodes, the amount of data transmitted between nodes is currently the root cause of the limitations on the performance and bandwidth transmission of these systems.

Because some nodes in the system are more than one kilometer apart, and metal wires dissipate electrical signals in the form of heat when transmitting data at high speeds, people usually use optical fibers to transmit data. But a new problem that arises from this is that when signals are sent from one node to another, there is also a significant amount of energy loss in the mutual conversion process between electrical and optical signals.

Based on this, a research team from the School of Engineering at Columbia University in the United States combined wavelength division multiplexing technology and Kerr frequency comb to design a simple and energy-saving data transmission method. Compared to the traditional method of transmitting multiple signals simultaneously using the same optical cable, this new technology does not require the use of different lasers to generate light of different wavelengths, but only requires one laser to generate hundreds of different wavelengths of light, which can simultaneously transmit independent data streams. This achievement was published on Nature Photonics under the title of "Massively scalable Kerr comb driven silicon photonic link" (DOI: 10.1038/s41566-023-01244-7).

The Kerr frequency comb can utilize the nonlinear optical Kerr effect in a micro resonant cavity to transform a single frequency laser into a broadband optical frequency comb containing a large number of equally spaced frequencies, and output an ultra short soliton pulse sequence in the time domain. Keren Bergman, a professor in the Department of Electrical Engineering at Columbia University's School of Engineering, stated that they can encode independent information channels for different frequencies of light through a Kerr frequency comb and transmit them through a single fiber optic. This breakthrough can enable the system to transmit more data without consuming more energy.

The research team integrates all optical components onto a millimeter scale chip to generate light, encodes it using electrical signals, and then converts the optical signals back into electrical signals at the target node. They designed a novel optical structure that can encode data separately for each channel while minimizing interference to adjacent channels. This means that signals sent with different wavelengths of light will not be confused and will not make it difficult for the receiver to demodulate.

Anthony Rizzo, the main member of the study, stated that their designed structure is more compact compared to other methods. This enables this type of chip to be directly connected to computer electronic chips, and the transmission distance of electrical signals has changed from a few tens of centimeters to a few millimeters, greatly reducing total energy consumption. Moreover, frequency combs based on silicon nitride can be manufactured in standard CMOS foundries for processing microelectronic chips, without the need for expensive specialized III-V foundries, resulting in lower costs.

Bergman pointed out that this study provides a feasible path for significantly reducing system energy consumption while improving computing power, and is expected to enable artificial intelligence applications to continue growing at an exponential rate. In the experiment, researchers successfully transmitted 32 different wavelengths of light at a speed of 16 GB per second, with a total bandwidth of 512 Gb/s for a single fiber. And in the data of one trillion transmission bits, the error is less than one bit. The size of the silicon chip used to transmit data is only 4 passwords x 1 mm, while the size of the chip used to receive optical signals and convert them into electrical signals is only 3 mm x 1 mm, both of which are smaller than human nails.

Figure Classification Data Center of Silicon Photon Link Driven by Kerr Frequency Comb


Although we used 32 wavelength channels in the principle verification, our architecture can be expanded to accommodate over 100 channels, which is completely within the range of standard Kerr frequency comb design, "Rizzo added. Importantly, these chips can be manufactured using the same devices as microelectronic chips used in standard consumer laptops or mobile phones, providing a direct path for volume expansion and practical deployment. The team's next step is to integrate photonics with chip level driver and control electronics to further miniaturize the system.

Source: Chinese Journal of Light Sources


Related Recommendations
  • Overview of Ultra Short Pulse Laser Processing of Wide Bandgap Semiconductor Materials

    Professor Zhang Peilei's team from Shanghai University of Engineering and Technology, in collaboration with the research team from Warwick University and Autuch (Shanghai) Laser Technology Co., Ltd., published a review paper titled "A review of ultra shot pulse laser micromachining of wide bandgap semiconductor materials: SiC and GaN" in the international journal Materials Science in Semiconductor...

    2024-07-30
    See translation
  • Scientists demonstrate a new optical neural network training method that can crush electronic microprocessors

    The current deep neural network system (such as ChatGPT) can quickly improve energy efficiency by 100 times in training, and "future improvements will greatly increase by several orders of magnitude. Scientists from MIT and other institutions have demonstrated a new optical neural network training method that can crush state-of-the-art electronic microprocessors.Moreover, the computational density...

    2023-09-27
    See translation
  • Devices based on optical thermodynamics can guide light without the need for switches

    Researchers from Ming Hsieh's Department of Electrical and Computer Engineering at the University of Southern California have designed the first optical device that follows the emerging optical thermodynamic framework.The work, reported in Nature Photonics, introduces a new way of routing light in nonlinear systems—meaning systems that do not require switches, external control, or digital addressi...

    10-15
    See translation
  • Ireland's first biological Brillouin microscope at Trinity College Dublin

    A project at Trinity College Dublin is now hosting Ireland's first BioBrillouin microscope instrument, applying Brillouin spectroscopy to life sciences and medicine.This should in particular enhance the College's research into cellular and tissue mechanics for the study of inflammation, cancer, and developmental biology.Brillouin microscopy offers a route to optical investigation of a biological s...

    07-14
    See translation
  • Germany's leading optoelectronics industry (Jenoptik) in the first half of the gold over 4.2 billion

    On August 9, local time, Germany's leading optoelectronics company Jenoptik released its 2024 second quarter interim financial results forecast. The financial data show that the company in the challenging market environment still shows strong growth momentum.In the first half of the year, Jenoptik achieved significant growth in revenue and earnings before interest, taxes, depreciation and amortiza...

    2024-08-15
    See translation