Kneron, the San Diego- and Taiwan-based AI silicon and IP startup, has launched an AI SoC which features an updated version of the company’s neural processing unit (NPU) IP. The KL720 also features a Cadence DSP AI co-processor and an Arm Cortex M4 core for system control. While Kneron’s next-gen AI SoC is aimed at low-power edge and smart home devices such as video doorbells and robot vacuum cleaners, the KL720 “can be used in everything from a Tesla to a toaster,” according the company.
Kneron claims this second-generation chip outperforms chips from both Intel’s Movidius line and Google’s Coral Edge TPU in terms of energy efficiency. The KL720’s NPU block can perform 1.4 TOPS while the whole SoC, including the additional DSP and Cortex M4 cores, comes in at 0.9 TOPS/W. This is sufficient for processing 4K resolution images and videos up to Full HD 1080p resolution. This compares favorably to Kneron’s previous generation chip, KL520 which was released in May 2019, which could achieve 0.3 TOPS at 0.6 TOPS/W.
While the previous generation chip was aimed solely at image processing, Kneron’s next-gen AI SoC is also a good fit for audio processing. With the increasing popularity of voice control interfaces, there is increasing demand for AI processing inside the edge device since it is quicker and cheaper than processing in the cloud, and maintains user privacy. Kneron says the KL720 has enough processing power to recognize “an entire dictionary worth of words,” far beyond competing chips which can recognize specific wake words only.
Kneron has been demonstrating a prototype of the KL720 to customers since at least January. Founded in 2015, the company started out developing AI models for use cases including facial recognition. As well as AI silicon, the company licenses its NPU IP; the version of the NPU in the KL720 has already been successfully integrated with Cadence Tensilica Vision P6 DSP IP and Synopsys’ ARC processor IP.
The key that allows the NPU to work with both images and audio is its reconfigurable design.
“We break down mainstream AI frameworks and [convolutional neural network] 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,” Kneron CEO Albert Liu told EE Times in an earlier interview.
“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.”
Kneron also recently announced Kneo, the company’s private mesh network for connected AI-powered sensors. Kneo is designed to allow a consumer’s devices (at least, those that include a Kneron chip) to work together without sending any data to the cloud. Data is instead stored locally, protected by blockchain. The company said Kneo will also allow consumers to keep their data away from “big tech,” and even sell their own data on their own terms, if they wish.
Samples of the KL720 will be available “soon”.
>> This article was originally published on our sister site, EE Times.