Syntiant, a US based AI chip start-up, announced the launch of low-power neural network accelerators ndp100 and ndp101 designed for speech task processing, which will be used to detect sound modes with power levels below 200 μ W, so that various devices can have voice control functions.
What’s amazing about the new chip is the use of digital technology. This start-up company made its debut in 2018 and introduced a method of using analogy technology to process deep learning tasks by using an array of multiply accumulate units containing hundreds of thousands of units connected to nor memory. Rival rival rival rival mythic specializes in imaging and video applications with similar analog technology, while syntiant said that the next generation chip to be launched in 2020 will also lock in such applications.
The architecture of processor in memory (PIM) has always been regarded as an interesting but difficult method to implement. The two start-ups, rival mythic and syntiant, may need to move from 40 nm nor to ReRAM or MRAM array to design in the process of miniaturization to 28 nm.
The new chip uses the SRAM cache memory in the digital MAC array. Part of the way to save power is to reduce the data movement as much as possible, and reduce the accuracy level. It supports 4-bit weights and 8-bit activations. The introduction of digital components represents the company’s awareness of the importance of neural network pruning and the challenges brought about by analog computing.
Syntiant claims that its new digital chips are better than ordinary CPUs and chips DSP can save 100 times of energy consumption, but it does not provide relevant measurement data. The company said that its digital chip has won the design of hearing aids, mobile phones and other products, and the chip samples have been tested in smart speakers, home automation devices and laptops, while the analog chip is still in the research and development stage.
A senior executive of Motorola solutions was quoted as saying that the new digital chips could enable public security systems to realize “a new generation of edge computing applications”; another senior executive of Infineon said that without DSP or cloud connection, a single im69d130 would be able to realize “a new generation of edge computing applications” The microphone works well in the near end and far end audio processing.
Syntiant said the digital chip will be used for keyword spotting, wake word detection, speaker recognition, speech event recognition and sensor analysis and processing. This product supports 63 spoken words, and can recognize sounds through programming, such as broken glass sound. In addition, it can monitor the data of gas sensors, or use passive infrared data for human body detection.
The start-up plans to use analog memory processors to replace the digital blocks of the new chip (pink in the picture) in the future to save more power without changing the surrounding IP blocks (blue in the picture).
The company integrates a utility into tensorflow architecture to automate the quantification process, and embeds the deep learning model into the chip to support other architectures in the future. Kurt Busch, executive director of syndiant, said: “we provide development tool kits (SDKs) and training development kits that can interface with other processors for customers who want to train their own models. We also provide voice training services for customers who want to customize keywords.”
The NDP 100 is 1.4×1.8 mm in size and is packaged with 12 ball wlbga, which can be placed in mobile phones and hearing aids with limited space. NDP 101 is packaged in 5 × 5mm QFN and can be started from serial flash memory. It has 8 gpios and can be used as the main system chip in large-scale systems such as smart speakers. Both products are equipped with arm Cortex-M0 core with 112-kb ram, and have begun to ship.
Editor in charge: PJ