Perceive’s AI processor; tactile sensors; Tbit/s on a twisted pair; FPGA acceleration of legacy programs

A new edge inference solutions company has been making news over the past few days, claiming an outstanding power efficiency. Other updates this week are mostly from research works in different areas.

AI edge processing at 55 TOPS/Watt

On March 31st Perceive Corporation emerged from stealth mode and debuted its first product, the Ergo edge inference processor. According to Perceive, Ergo delivers more than 4 sustained “GPU-equivalent” floating-point TOPS, with the ability to run heterogeneous, large neural networks simultaneously, and offering a power efficiency of more than 55 TOPS/Watt. For example, Ergo can run YOLOv3 at up to 246 frames per second (batch size =1) at 30 frames per second while consuming about 20 mW. The processor has a 7×7 mm package and requires no external RAM. Ergo targets applications such as video object detection, audio event detection, and speech recognition, in consumer devices such as security cameras, smart appliances, and mobile phones. Perceive also announced that Ergo has been selected by two major providers of smart connected camera and security products – one of them being Arlo – to integrate advanced neural network applications into future products. The Ergo chip and reference board are currently being sampled to leading customers; the company expects to be ready for mass production in the second quarter of 2020. Founded in 2018 and based in San Jose, CA, Perceive is a majority-owned subsidiary of Xperi Corporation, the company known for brands such as DTS, IMAX Enhanced, HD Radio, and Invensas.

Neural networks improve phone call quality

Google’s Duo calls on Pixel 4 phones can now benefit from a new “packet loss concealment” technique based on neural networks. As Google researchers explain in this blog post, the new solution used to fill the voice gaps caused by lost data packets is a modified version of WaveRNN, a recurrent neural network model for speech synthesis consisting of two parts: an autoregressive network and a conditioning network. The autoregressive network is responsible for the continuity of the signal and provides the short-term and mid-term structure for the speech by having each generated sample depend on the network’s previous outputs. The conditioning network influences the autoregressive network to produce audio that is consistent with the more slowly-moving input features. The result is a better audio quality.

Vision-based tactile sensors for robotics

Researchers around the world are trying to overcome the limitations of existing tactile sensors for robotics applications, that are either flat, small or low-resolution. A team of researchers at UC Berkeley has recently develop OmniTact, a multi-directional high-resolution tactile sensor designed to be used as a fingertip for robotic hands. The device uses multiple micro-cameras to detect multi-directional deformations of a gel-based skin. This provides a rich signal from which a variety of different contact state variables can be inferred using image processing and computer vision methods – including neural networks. Researchers found that OmniTact can successfully perform challenging robotic control tasks, such as reliably inserting an electrical connector into an outlet. This work is set to be presented at ICRA 2020 (International Conference on Robotics and Automation).

A vision-based approach to tactile sensors is also being pursued by ETH Zurich (Switzerland), where researchers have developed an inexpensive device consisting of an elastic silicone “skin” containing colored plastic microbeads, and a regular camera affixed to the underside. When the sensor comes into contact with an object, an indentation appears in the silicone skin. This changes the pattern of the microbeads, which is detected by the fisheye lens on the underside of the sensor. Analyzing pattern changes allows to calculate the force distribution on the sensor.

Terabit-per-second on a twisted pair

Using the same principle on which Digital Subscriber Line is based, a team of researchers has transmitted signals at 10 terabits per second over a distance of 3 meters, through a setup that emulated the metal-sheathed twisted pairs of telephone cables typically used for DSL service. Reportedly, data rate drops to 30 gigabits per second at a range of 15 meters. The research work employed a 200-gigahertz signal with the same vector coding used at lower frequencies in regular DSL systems, creating a solution where the twisted pair with its metal sheath work as a multi-mode waveguide, with advantages equivalent to a MIMO configuration in wireless systems. The research team includes John Cioffi, credited to be “the father of DSL,” now chairman and CEO of Internet connectivity firm Assia in Redwood City, CA. This technique might be useful in applications that require high transmission rates over short distances, such as between racks in a data center or for chip-to-chip connections.

John Cioffi. Image credit: Assia

Automatic conversion of legacy code for FPGA acceleration

Using FPGAs to accelerate CPU legacy programs often requires code rewriting at very low levels of abstraction and an expert’s knowledge of the target accelerator architecture. Now a new approach has been proposed by a team from University of Michigan. The technique combines dynamic and static analyses to learn a model of functional behavior for off-the shelf legacy code, then synthesizes a hardware description from this model. The team developed a framework that transforms Boolean string kernels into hardware descriptions using techniques from both learning theory and software verification. In a paper presented at the ASPLOS conference in Lausanne, Switzerland, the team demonstrated their technique on one class of functions, string kernels, used for search and comparison on text. Application of this approach to other classes of functions will be the subject of future investigations.

Acquisitions

Diodes Incorporated has received all required regulatory approvals from the Taiwan authorities for its proposed acquisition of Lite-On Semiconductor. Diodes is now working to comply with regulatory procedures in China and remains confident the transaction will close as planned. Diode product portfolio includes discrete, analog, and mixed-signal products with advanced packaging technology. Lite-On specializes in bridge rectifiers and other power-related components.

MagnaChip Semiconductor has announced a definitive agreement to sell its foundry business and Fab 4. According to the company, this move will streamline MagnaChip into a pure-play standard products company, positioning it to focus on high growth markets in analog power and display solutions including OLED and microLED. MagnaChip foundry business and Fab 4 will be acquired by a “special purpose company” in South Korea established by Korean investors, with other limited partners including SK hynix.

This entry was posted on Friday, April 3rd, 2020 at 4:02 pm. You can follow any responses to this entry through the RSS 2.0 feed. You can leave a response, or trackback from your own site.