Syntiant, an AI chip start-up based in Irvine, California, received the news on August 7 that its Neural Decision Processors (NDP), NDP100 and NDP 101, have been certified by Amazon. This allows the integration of the chips into a variety of Alexa-based devices.
The company is already well supported by a number of heavyweight tech investors including Amazon’s Alexa Fund, Intel, Microsoft’s M12 Venture Fund and Robert Bosch Venture Capital GmbH. To date, Syntiant has raised US$30 million.
Clearly, being certified by Amazon is another major coup for the company and will enable device OEMs to offer a built-in “always-on” Alexa listening experience in even the most power-sensitive battery-powered devices including earbuds and Bluetooth headsets, smartwatches, IoT sensors, and remote control devices as well as introducing voice control into other entirely new form-factor devices.
This will considerably expand the Amazon Echo ecosystem and offers a potentially huge market opportunity for Syntiant. Counterpoint forecasts that the wireless hearables market alone will grow to 175 million units globally by 2021.
Interestingly, Syntiant already has a partnership with earbud vendor Bragi, who is using Syntiant’s processors in its wireless earbuds and in May it announced that it is also working with Taiwanese gaming company MSI to introduce “voice in-game” features via Amazon Alexa.
Let us take a closer look at the newly certified Syntiant products. The NDP10x series of speech and audio processors are custom built to run neural workloads. They are primarily designed for integration into various types of voice and audio-enabled devices. According to Syntiant, the processors can recognize up to 64 words or other sensor patterns while consuming just 150 microwatts. The company claims that this is a 200-fold improvement over what a typical microcontroller can offer.
With dimensions of just 1.4 mm x 1.8 mm, the chip is supplied in a 12 ball WLBGA package and is typically connected directly to a digital microphone which triggers a larger, usually sleeping, system within a device (Exhibit 1). Once awake, this system interrogates the NDP100 to determine which wake word or command it heard. The chip also has a three-second audio buffer in case the system needs to catch up on what was said during its wakeup routine.
As with most AI ASICs, the chip only performs inferencing. Training happens in the cloud using Google’s TensorFlow software library with the resultant neural network parameters programmed directly into the chip as firmware using Syntiant’s proprietary algorithms. The inference engine can classify 100 words per second.
Exhibit 1: Syntiant’s NDP100 voice-recognition chip
Syntiant’s technology is based on “processor-in-memory” architecture, and it is one of a small band of start-ups focused on developing an all-analog memory computing solution. However, its first products, the NDP10x series of processors use digital multiply-accumulate (MAC) units rather than flash memory-based multipliers.
Performing computation “in-memory” eliminates both the memory bandwidth and memory power penalties normally associated with CPU-based processors, resulting in significant reductions in overall power consumption. Other start-us involved in this space include Mythic and Gyrfalcon, but Syntiant seems to be the only company focused on low-power audio applications.
In future products, the company intends to replace the digital MACs with low precision but very accurate analog MACs which should result in further, and possibly significant, power savings. This will make its chips even better suited to ultra-low-power applications. However, mass-producing an analog-based design presents its own set of challenges, which is perhaps why Syntiant initially opted for a digital MAC design.
The company is also developing a 20 TOPS/Watt NPD chip to process video which should be sampling in H2 2019 and should enable it to expand into more markets.
Exhibit 2: Syntiant NDP 100 architecture
Syntiant is one of ten AI chip start-ups profiled in Counterpoint Research’s upcoming report on the AI chip start-up market. Other companies profiled include Efinix, FlexLogix, Graphcore, Gyrfalcon, Habana Labs, Mythic, ThinCI and Wave Computing.
