Power/Performance Bits: Feb. 23

Photonic AI accelerator; a clearer look at HDD write heads; drawing circuits with a pen.


Photonic AI accelerator
There are now many processors and accelerators focused on speeding up neural network performance, but researchers at the University of Münster, University of Oxford, Swiss Federal Institute of Technology Lausanne (EPFL), IBM Research Europe, and University of Exeter say AI processing could happen even faster with the use of photonic tensor processors that can handle multiple data sets at once.

The team created a photonic hardware accelerator for matrix multiplications, the main processing load in neural networks. They combined this with phase-change materials (PCMs) as energy-efficient storage elements. Commonly used in optical data storage, PCMs allow the new processor to store and preserve the matrix elements without the need for an energy supply.

To carry out matrix multiplications on multiple data sets in parallel, the physicists used a chip-based frequency comb as a light source. A frequency comb provides a variety of optical wavelengths which are processed independently of one another in the same photonic chip. This enables highly parallel data processing by calculating on all wavelengths simultaneously. “Our study is the first one to apply frequency combs in the field of artificial neural networks,” said Wolfram Pernice, a professor from the Institute of Physics and the Center for Soft Nanoscience at the University of Münster.

A convolutional neural network for recognizing handwritten letters was used in testing, but the researchers said it could apply to a range of convolution-based tasks. Plus, they noted that the approach provides a path to fill CMOS wafer-scale integration of the photonic core.

“The convolutional operation between input data and one or more filters – which can be a highlighting of edges in a photo, for example – can be transferred very well to our matrix architecture,” explained Johannes Feldmann, now a postdoctoral researcher at University of Oxford. “Exploiting light for signal transference enables the processor to perform parallel data processing through wavelength multiplexing, which leads to a higher computing density and many matrix multiplications being carried out in just one timestep. In contrast to traditional electronics, which usually work in the low GHz range, optical modulation speeds can be achieved with speeds up to the 50 to 100 GHz range.”

A clearer look at HDD write heads
As the volume of data generated increases every year, hard disk drives (HDDs) remain the primary data storage device with the $20 billion worth of HDDs shipped in 2020 expected to have a total capacity of over one zettabyte. However, understanding write head operations poses a barrier to improving capacity and data transfer rates.

Current write heads have a very fine structure, with dimensions of less than 100 nanometers. Magnetization reversal occurs in less than a nanosecond, making experimental observations of write head dynamics difficult.

Researchers from Tohoku University, Toshiba Corporation, and the Japan Synchrotron Radiation Research Institute (JASRI) used the large-scale synchrotron radiation facility SPring-8 to image the magnetization dynamics of an HDD write head for the first time, with a precision of one ten-billionth of a second.

To accomplish this, a write head is operated at an interval of one-tenth of the cycle of the periodic X-ray pulses generated from the SPring-8 storage ring. Simultaneously, focused X-rays scan the medium-facing surface of a write head, and magnetic circular dichroism images temporal changes in the magnetization. This achieves temporal resolution of 50 picoseconds and spatial resolution of 100nm, enabling analyses of the fine structures and fast write head operation. The researchers said the method has the potential to achieve even higher resolutions by improving the focusing optics for the X-rays.

With the imaging, the team found that magnetization reversal of the main pole is completed within a nanosecond and that spatial patterns from magnetization appear in the shield area in response to the main pole reversal.

They expect the approach to support high-precision analyses of write head operations, contributing to the development of the next-generation write heads and the further improvements in HDD performance. Toshiba aims to apply the developed analysis method and the knowledge obtained about write head operations to the development of a write head for energy-assisted magnetic recording.

Drawing circuits with a pen
Conductive inks are used to print or draw flexible circuits on surfaces. However, they can be expensive, clog the application device, and fail on some materials.

Researchers from Wuhan University developed an inexpensive conductive ink and a clog-free ballpoint pen that can be used to draw circuits almost anywhere.

Clogging tips are a particular problem for pen-style applicators of conductive inks. Instead of the more expensive metal-particle based inks, the team turned to a water-based ink containing conductive carbon particles composed of graphene nanosheets, multiwalled carbon nanotubes and carbon black. Maleic anhydride modified rosin resin was added as a binder to reduce the ink’s solid content and viscosity, and xanthan gum was added to stabilize the dispersion so the carbon wouldn’t settle out of the ink.

The researchers optimized viscosity and the size of the conductive particles relative to the pen tip to create a system that provided stable and smooth writing performance on both flat and irregular surfaces. They demonstrated the ink’s ability to work on difficult surfaces by drawing circuits on a loofah.

Circuits drawn on paper with the pen withstood multiple cycles of folding without deterioration. The ink remained stable after sitting for 12 hours, released no harmful gases during use and cost much less than others reported in the literature, the researchers noted. The pens could also be used to draw flexible, wearable electronic devices on soft substrates or human skin.

Leave a Reply

(Note: This name will be displayed publicly)