Designing an AI inference engine that can power neural networks including speech recognition, enhance low light photo or any other image recognition techniques takes time.
Gyrfalcon technology took time to manufacture two SoC based accelerators called Lightspeeur® 2801 and 2803 and offered them to the customers. Now after the initial SoC business offering it realized that some customers would want to purchase the IP rights and implement it in their solutions. This is a widely accepted business model in the SoC manufacturing as you cannot possibly invent all the IP in house and it makes more financial and time to market constraints to acquire and implant other’s IP.
The accelerator chips’ architecture features AI cores that are ultra-small and low-power, enabling AI Processing in Memory (APiM). The AI cores configured in a proprietary Matrix Processing Engine (MPETM) architecture. The AI cores accelerate the convolutional neural network (CNN) on AI frameworks like Caffe, PyTorch, and TensorFlow.
16.8 TOPS under one W
GTI’s accelerator chips combine over 28,000 cores, with the Lightspeeur 2803 capable of 16.8 TOPS in under a watt, and the Lightspeeur 2801 using only 300mW while providing 2.8 TOPS. This is a great achievement and having a custom designed SoC to run CNN’s is better than a generalized solution where one fits many.
Offering IP to third parties can help the SoC designers with faster integration of an AI inference acceleration at a minimal cost in die area and just milliwatts of power consumption.
“GTI is differentiating itself at a time when the market has been saturated with companies talking about AI chips. We have been producing our three different AI Accelerator chips, now in the hands of customers designing end products, so our silicon is proven. While we intend to continue introducing and producing new AI Accelerator chips, our technology is 100% proprietary, and protected with a broad and deep portfolio of patents. We will deliver the advantages of GTI technology to companies seeking new advantages, accelerating the growth of both the global AI chip market and the monetization of our technology”, said Kimble Dong, CEO of Gyrfalcon Technology, Inc.
The company introduced its first Lightspeeur 2801 chip as an Edge AI solution in 2017, and added a second chip, the Lightspeeur 2803 optimized for cloud AI in the data center in 2018. Both are offered in a configuration of 16 chips on a PCIe Card called GAINBOARDTM System products, packaged for upgrading data centers needing to provide AI services.
Customers are designing Gyrfalcon chips inside numerous commercial products, including; smart home & office products, consumer & electronics, mobile phones & computers, baby and pet monitors, robot vacuums, defect detection equipment, edge servers and AI Data Center solutions.