Gyrfalcon Technology was founded in January 2017 to develop low-cost, low-power, high-performance Arti­ficial Intelligence (AI) processors. Founded by veteran Silicon Valley entrepreneurs, Gyrfalcon aims “to expand the power of cloud Artificial Intelligence to local devices with greater performance and efficiency.”

The company’s first AI chip, the Lightspeeur 2801S Neural Pro­cessor, features high performance and superior energy efficiency (5.6 TOPS/Watt), enabling AI Edge Computing and Datacenter Machine Learning up to several orders of magnitude faster than MCUs and GPUs. The device is focused on infer­ence on the edge, which Gyrfalcon believes will account for most of AI computing. Training will still rely on Nvidia in the cloud. The device is an AI accelerator and is designed to serve as a co-processor to a host CPU. In the future, the company will consider developing a complete SoC with integrated AI engine and host CPU.

Lightspeeur is based on Gyrfalcon’s APiM architecture, which uses mem­ory as the AI processing unit. This eliminates the huge data movement in other architectures, which results in high power consumption. The architecture features true, on-chip parallelism, in situ computing, and eliminates memory bottlenecks. The device has roughly 28K paral­lel computing cores and does not require external memory for AI inference.

Lightspeeur supports neural net­works such as Convolutional Neural Networks (CNN), ResNet, RNN and LSTM. The distributed memory block structure is friendly to CNN computing and supports multiple layers and different sizes for each layer. The built-in model compress algorithm enables fast and low power CNN computing. It supports standard, open frameworks such as Caffe, TensorFlow, and MXNet.

Software Development Kits include turnkey designs, system verifica­tion hardware, software and tools. As of this writing, Gyrfalcon has ported their software onto their SDK, with features including Fast Image Recognition, Pixel-level Im­age Segmentation, Real-time Video Analysis, Speech Recognition and Translation, and Handwriting Chi­nese Character Recognition.

The LightSpeeur 2801S was deliv­ered from TSMC on Sep 19, 2017, fabricated on a 28nm process. With a VGG type neural network, fully running 142 frames/sec­ond (224*224*3) pictures, power consumption is 0.3W at 50MHz working clock. VGG is a CNN model from the University of Oxford that achieves high accuracy. With partial­ly running moving video recognition frame by frame, 32 frames/second (224*224*3) pictures, power con­sumption is 0.2-0.3W at 50MHz or less working clock.

Numerous companies, both large and small, are developing AI pro­cessors. Gyrfalcon believe it has at least a year head start, with silicon in hand now. LightSpeeur is based on Dr. Yang’s 30 years of neural network research and features true digital neural network circuit­ry. Gyrfalcon argues that many of today’s competitors only simulate neural networks.

In the Cloud, the device can be used in AI inference servers. On the Edge, the device can enable AI powered surveillance/video cams, smart toys/robotics, smart home devices, AR/VR products, speech/ voice recognition, natural language processing and more. Gyrfalcon is in active discussions with domestic AI institutions and universities as well as overseas entities. Several tier-one companies plan to design Light­speeur into applications including servers, cell phones, smart city products, and surveillance cameras.