| IP개요 |
As artificial intelligence (AI) applications demand significant resources, hardware and software optimization for target applications has become a crucial element. Especially, in embedded systems where resource utilization is limited, efforts to minimize hardware utilization for AI applications have emerged as an important issue. Applying simpler AI algorithms is suggested as one alternative for this. Meanwhile, AI applications are diverse, including image-based and sound-based applications, therefore, the details of hardware or software vary depending on the application. While software updates in real-time, the static characteristics of hardware architecture make it hard to reconfigure. In this work, we propose Intellino, an AI processor that performs lightweight AI algorithms and reconfigure hardware according to the application while learning and inferring. The Intellino performs inference utilizing k-nearest neighbor (k-NN) or radial basis function (RBF) algorithms. Additionally, the Intellino allows reconfiguring hardware represented by the number of datasets and the individual size of the dataset within the memory size, enabling the exploring of optimized hardware for each AI application through experiments. The proposed AI processor is designed with Verilog HDL and verified through simulation. The designed architecture is implemented through a filed programmable gate array (FPGA) and its operation is verified through experiments for various applications. We will fabricate the Intellino with Samsung 28nm RFCMOS 1-poly 10LM technology. The designed processor operates in 1.2V, 50MHz frequency and the circuit is digital type. |
