Last week we briefly addressed the theme of machine learning in chip design; this week a Synopsys announcement provides a significant update on this topic. Other news includes some interesting academic research work.
Exploring design space with artificial intelligence
Synopsys has introduced DSO.ai (Design Space Optimization AI), what it claims to be the industry’s first autonomous artificial intelligence application for chip design, capable of searching for optimization targets in very large solution spaces. DSO.ai ingests large data streams generated by chip design tools and uses them to explore search spaces, observing how a design evolves over time and adjusting design choices, technology parameters, and workflows to guide the exploration process towards multi-dimensional optimization objectives. The new AI application uses machine-learning technology invented by Synopsys R&D to execute searches at massive scale: according to the company, DSO.ai autonomously operates tens-to-thousands of exploration vectors and ingests gigabytes of high-velocity design analysis data – all in real-time. At the same time, the solution automates less consequential decisions, like tuning tool settings. The announcement press release includes a quote from early-adopter Samsung, testifying that Synopsys’ DSO.ai systematically found optimal design solutions that exceeded previously power-performance-area results. Furthermore, DSO.ai was able to achieve these results in three days – as opposed as over a month of experimentation when the process is performed by a team of expert designers.
New Xilinx adaptive compute acceleration platform
Xilinx has announced Versal Premium, the third series in the Versal ACAP (adaptive compute acceleration platform) portfolio. The Versal Premium series is built on a foundation of the currently shipping Versal AI Core and Versal Prime ACAP series. New and unique to Versal Premium are 112Gbps PAM4 transceivers, multi-hundred gigabit Ethernet and Interlaken connectivity, high-speed cryptography, and PCIe Gen5 with built-in DMA, supporting both CCIX and CXL. The new platform is designed for the highest bandwidth networks operating in thermally and spatially constrained environments, as well as for cloud providers who need scalable, adaptable application acceleration.
Image credit: Xilinx
Low power audio AI applications at the edge
Cadence has optimized the software of its Tensilica HiFi DSPs to efficiently execute TensorFlow Lite for Microcontrollers, part of the TensorFlow end-to-end open-source platform for machine learning from Google. This promotes rapid development of edge applications that use artificial intelligence and machine learning,removing the need for hand-coding the neural networks. According to Cadence, Tensilica HiFi DSPs are the most widely licensed DSPs for audio, voice and AI speech; support for TensorFlow Lite for Microcontrollers enables licensees to innovate with ML applications like keyword detection, audio scene detection, noise reduction and voice recognition, with an extremely low-power footprint.
Image sensor with neural network capability
Researchers from Vienna University of Technology (Vienna, Austria) have demonstrated how an image sensor can itself constitute a neural network that can simultaneously sense and process optical images without latency. The device is based on a reconfigurable two-dimensional array of tungsten diselenide photodiodes, and the synaptic weights of the network are stored in a continuously tunable photoresponsivity matrix. In other words, the sensitivity of each photodiode can be individually adjusted by altering an applied voltage, and sensitivity factors work like weights in a neural network. By creating the appropriate sensitivity pattern, the image sensor as a whole acquires the ability to perform some basic machine learning function. The experimental device is a square array of nine pixels, with each pixel consisting of three photodiodes; resulting currents (analog signals) are summed along a row or column, according to Kirchhoff’s law. The researchers demonstrated that the device could sort an image into one of three classes that correspond to three simplified letters, and thus identify which letter it is in nanoseconds. Throughput is in the range of 20 million bins per second. Practical applications of this interesting concept would require solving a number of problems inherent to the technology used in the research chip, such as difficult imaging under dim light, high power consumption, difficult manufacturing over large areas etc. With different sensors, the same concept could be extended to other physical inputs for auditory, tactile, thermal or olfactory sensing.
Image credit: Nature
Nanoelectromechanical non-volatile memories withstand 200°C
Nanoelectromechanical relays have emerged as a promising alternative to transistors for creating non-volatile memories that can operate in extreme temperatures with high energy efficiency. However, until now, a reliable and scalable non-volatile relay that retains its state when powered off has not been demonstrated. Now researchers from University of Bristol, in collaboration with the University of Southampton and the Royal Institute of Technology (Sweden), have come up with a new architecture that overcomes those limitations. As the team explained, part of the challenge is the way electromechanical relays operate. When actuated, a beam anchored at one end moves under an electrostatic force; as the beam moves, the airgap between the actuation electrode and beam rapidly reduces while the capacitance increases. At a critical voltage called the pull-in voltage, the electrostatic force becomes much greater than the opposing spring force and the beam snaps in. The Bristol team explained that this inherent electromechanical pull-in instability makes precise control of the moving beam, critical for non-volatile operation, very difficult. The new device, instead, is a rotational relay that maintains a constant airgap as the beam moves, eliminating this electromechanical pull-in instability. Using this relay, the researchers have succeeded in demonstrating the first high-temperature non-volatile nanoelectromechanical relay operation, at 200 °C. Potential applications include electric vehicles as well as zero-standby power intelligent nodes for the IoT.
Image credit: Dr Dinesh Pamunuwa
Acquisitions
Ansys has entered into a definitive agreement to acquire Lumerical, a developer of photonic design and simulation tools. With optical networks becoming increasingly important in data center architectures and other applications, Lumerical’s products enable designers to model problems in photonics, including interacting optical, electrical and thermal effects. Infineon Technologies will proceed to acquire Cypress Semiconductor; the Committee on Foreign Investment in the United States (CFIUS) has concluded its review of the planned acquisition and cleared the transaction. TE Connectivity, a provider of connectivity and sensing solutions, completed its public takeover of First Sensor, a German player in sensor technology. TE now holds 71.87% shares of First Sensor. Silicon Labs has entered into a definitive agreement with Redpine Signals to acquire the company’s Wi-Fi and Bluetooth business, development center in Hyderabad, India, and patent portfolio for $308 million in cash. The integration of the Redpine Signals technology is expected to accelerate Silicon Labs’ roadmap for Wi-Fi 6 silicon, software and solutions.
