Syntiant raises a series C round, bringing total funding to $65m. The company has shipped a million of its first gen part, but has second and third gen parts already in the pipeline.
Ultra-low-power AI accelerator startup Syntiant has raised another $35 million in a series C round of funding to bring the total raised by the company to $65 million. Syntiant, whose 66 staff work out of Irvine, Calif., also announced that it has hit a shipping milestone with 1 million parts in the hands of customers.
Syntiant’s C round was led by Microsoft’s VC fund, M12, and Applied Ventures, the VC arm of Applied Materials.
“[$35m] gets us pretty far into growing our sales team and ramping our revenue,” Syntiant CEO Kurt Busch told EE Times. “We have the second-generation chip already back in the lab, which we expect to announce before the end of the year… this funding will also be used to fund development of third generation silicon and build out our customer base.”
While Syntiant’s first-gen chip is suitable for sensor processing and voice processing (wake word detection, command detection and speaker ID), the second gen will add expand capabilities for both audio and image processing, and Syntiant is aiming for understanding conversational speech and more advanced context awareness functionality with its third gen. The company also develops neural networks that optimizes for its hardware and data training pipelines.
“The state of ML today is there are few companies who can take raw silicon and do stuff with it – these tend to be the industry whales – Google or Amazon,” Busch said. “To make mass market silicon useable, you need to have the training pipeline and the data because those are not readily commercially available to most companies.”
Syntiant’s offering is therefore positioned to be a turnkey solution, supplied together with trained neural networks for the specific application.
Hitting a million
The one million parts shipped to date includes both NDP100 and NDP101 parts since the company’s first production orders in September 2019. Both are manufactured at UMC in Singapore. The NDP100 is 1.4 x 1.8mm while the NDP 101 is slightly larger and includes more I/O options.
The parts are primarily used for wake word detection and speaker ID. Keyword detection can be done using the NDP100 with Syntiant’s neural network with just 3.4 µJ of energy at a maximum throughput of 100 audio frames per second. Syntiant’s figures have an Arm-M4 based microcontroller using 658 µJ at maximum 5 fps for the same task, though a different neural network is used.
Most of the million units sold have gone into laptops and cell phones, according to Busch, though he declined to name any customers. While most of the parts shipped so far are listening for speech, Busch said the company also has design wins in systems that use accelerometers or gas sensors, where Syntiant also develops neural networks for anomaly detection.
Production of Syntiant’s second generation device is on track for Q4 2020, said Busch.