Low-power deep learning accelerator integrates neural network design / training, model compressionaccelerator design. It uses a small amount of computationmemory footprint to realize high performance on the edge device. The image semantic segmentation technology can reach 80 frames per second (resolution:1024*2048) to meet real-time requirements for autonomous driving.

Assisted by in-house AI deep compression toolchain (ezLabel, ezModel, ezQUANT, ezHybrid-M), the proposed technology supports automatic AI model designoptimization with the integrated performance of 120x model size reduction70x power reduction in 2D CNN model,develops a world-first 1/2/4/8-bit CNN model realized by the developed high efficiency Hybrid fixed point CNN NPU (Hybrid-NPU), which has been verified in Xilinx ZCU102 FPGAachieves the performance up to 2.5 TOPS(8-b)/ 20TOPS(1-b)@28nm technology running at 550MHz4TOPS/W energy efficiency.

Electroplated nanotwinned Cu possesses excellent electrical & mechanical properties. It can be applied in three major joints: 1. Low thermal budget/ low resistance Cu bonding for high performance computing chip. We are able to achieve low temperature bondinginstant bonding. Low temperature bonding is performed at 150°C for 1 h to achieve low contact resistance copper bonding. Instant bonding is performed at 300°C for 5 seconds under a pressure of 90MPa achieve low contact resistance. 2. High strength/ High ductility copper lines in 3D-IC packaging We are able to fabricate high strength foils with tensile strength of 800MPa. After annealing at 150°C for 3 hours, the foil retains a tensile strength of 750MPa. 3. Cu foils for lithium ion battery

Dynamic vision sensors have been investigated to report motion-only images for moving object recognition. Less but essential information helps post-process recognition algorithm reduces computationimproves accuracy. Implement low powerlow latency deep Learning chip based on neuromorphic Intelligence. The neuromorphic obstacle detection algorithm integrates visualproprioceptive signals. The algorithm is characterized by its efficiencylow power consumption. We possess the next-generation in-memory computing AI chipsnext-generation UAV key softwarehardware technology