The company has not disclosed what manufacturing process technology the ASIC is being implemented in but describes the goal of the company as to combine the leading AI algorithms with advanced semiconductor technology.
ThinkForce was founded in 2017 and raised 450 million RMB (about $65 million) in a series A round of financing from YITU, the Jack Ma-backed YunFeng Capital, Sequoia Capital China and Hillhouse Capital.
The heart of ThinkForce's technology offering is neural network accelerator kernel named ManyCore that is architected to accelerate convolution, activation and pooling operations which are commonly used in NN inference, and support the combination of them. The external control unit can configure the kernel through asynchronous macro operation instructions. ManyCore uses data flow processing design without any instruction set.
As an FPGA-hosted prototype the architecture was based on 512 multiply accumulate units (MACs) running at 500MHz and was able to perform real-time human face detection for 16 HD video stream inputs. As yet there has been no disclosure about the data and weight resolution and the degree to which this can be controlled dynamically.
The chip is expected to also contain a video decoder to help address real-time analysis of digital video streams and offer a ten-fold increase in performance over the FPGA demo. The NNC-200V series cards, to come later, will be able to support face recognition and object classification for dozens of full HD video streams within a single PCIe card, the company said.
Related links and articles: