Second AI accelerator chip expected this summer.
Kneron, the San Diego and Taipei-based low-power edge AI startup, has raised $40 million in an A2 funding round. This funding round was led by previous backer Horizons Ventures and brings the company’s total financing to $73 million.
Kneron CEO Albert Liu told EETimes that the additional funding will be used for a variety of things, including support for the company’s upcoming second-generation AI processor, the KL720.
“[We plan to] hire engineers and developers to support current purchase orders, as well as forecasted purchase orders for the KL720, [and for] marketing and commercialization of our products, especially the KL720,” he said. He added that the company will also invest in “R&D of our future roadmap.”
The timeline for samples of the KL720 has slipped from Q1 2020 to mid-summer, Liu said, but he added that Kneron is already demonstrating a prototype of the new chip to customers. (EETimes notes that the first-gen chip, the KL520, was named after its launch date, May 20th — we therefore suspect the previously-unnamed KL720 might appear sometime around mid-July).
Hardware plus Algorithms
Liu believes Kneron’s integrated platform of hardware plus deep learning algorithms gives it a unique advantage. Founded in 2015, Kneron actually started out developing machine learning models for facial recognition, before quickly branching out into silicon. The company’s offering includes algorithms for facial recognition, face detection, body detection and gesture recognition. Its facial recognition model was recognized in 2019 by NIST as the best performing model under 100 MB (the model in question is 57 MB).
These models can be combined with the first-generation chip, the KL520, which is designed for mass-market edge devices. It is optimised for convolutional neural networks (CNNs) like those often found in image processing applications, and it can perform 0.3 TOPS at 0.5 W (equivalent to 0.6 TOPS/W).
Does this low-power chip pack enough punch to inference complex models like facial recognition?
“Yes,” said Liu, “because our MAC efficiency is over 90%, which is about three times the efficiency of competitors.”
Kneron also licenses IP for the architecture its chips are based on, a reconfigurable neural processing unit (NPU). This architecture can be reconfigured to switch in real-time between models, depending on the needs of the application.
“We break down mainstream AI frameworks and CNN models into basic building blocks and reconfigure them based on which application is needed and which AI framework we are working with so that our solutions can adapt to and accelerate the related CNN models,” Liu explained.
“For example, ResNet (for face recognition) and LSTN ( for voice recognition), though one is audio and the other is visual, have common building blocks,” Liu said. “While other solutions providers may need to support them with independent solutions, Kneron’s solution reconfigures the common building blocks in our reconfigurable AI engine so that in real-time, we can support different models like ResNet and LSTM based on the AI application.”
Successful partnerships have seen KL720 NPU IP integrated with Cadence Tensilica Vision P6 DSP IP and Synopsys’ ARC processor IP.
Company vs. its competition
As we begin to see some of the big players making AI-capable silicon for edge devices, the market is getting tougher and tougher. Liu welcomes this form of competition.
“Big players coming in further validates and pushes the edge AI space — more players, especially big ones will further push this space to grow and user adoption will increase,” he said. “Greater user adoption will raise market tides and lift all the boats in our space, so we welcome all players big and small to push edge AI along with us. At the end of the day, more competition will spur faster innovation and ultimately, consumers will be the winners.”
Liu also notes that edge AI is rather a wide spectrum.
“Some of the big players in edge AI don’t focus as much as we do on on-device edge, for example, we’re shipping with IoT products that need to focus on balancing power usage, performance, model size and cost,” he said.
The AI accelerator space is also crowded with dozens of other startups. Liu differentiates Kneron on several key criteria, including Kneron’s relatively advanced stage of development compared to some of its startup competitors. The company remains one of the earliest of the AI chip startups to get its product to the market, launching its flagship chip KL520 in May 2019. The company has 150 employees and is growing to keep up with market demand.
“We have a real chip and are already demoing a prototype of our second chip,” he said. “Also, you can already purchase some of our partners’ products on the market.”
In fact, Kneron has numerous customers already signed up. Aaeon has an accelerator module available that is built around the KL520, and it can also be found in IP security cameras from Brickcom and Amaryllo, smart door locks from Brickcom and DAAN (IDH), smart doorbells and intercoms from Primax and a smart door viewer from Xiaobei.
“We also have 40+ other purchase orders under NDA that we are working on,” Liu said.
KL720 silicon is expected to begin sampling this summer.