The international forum “AI in 5G – Leading the New Era Forum” was held in Shenzhen on March 6. At the conference, Artificial Intelligence Industry Alliance (AIIA) officially announced the first round of “AIIA DNN benchmark V0.5” evaluation results. As one of the products evaluated in the first round, Rockchip RK3399 passed the evaluation.
Mainly targeting the edge and based on AIIA’s authoritative test platform, AIIA DNN benchmark V0.5 objectively reflects the performance of processors with DNN acceleration capacity in completing inference tasks under four typical application scenarios. According to Sun Mingjun, the Leader of the General Group of AIIA, the first round of evaluation involved four typical application scenarios and two major categories of testing indicators, including speed (fps) and algorithm performance, such as top1, top5, mAP, mIoU, PSNR, etc.. It is the first benchmark in the field of deep learning processor that distinguishes integer and float comparison.
AI solution Rockchip RK3399 ,adoting 28nm process, participated in the first round of evaluation. The results showed that while fixed-point retraining was not required in float models, int8 achieved the performance double that of float computation at the price of a maximum precision loss of 1%.
Furthermore, in the Interpretation evaluation, AIIA tried to evaluate quantization and float models separately for the first time in the benchmark. The development board based on Rockchip RK3399, showed superior performance in a number of indicators, which were at the leading level in the industry.
The powerful performance of Rockchip RK3399 development board is also attributed to Tengine, a lightweight modular high-performance neural network inference engine developed by OPEN AI LAB. It specifically targets the optimization of Arm embedded devices without relying on third-party databases and supports Android and Liunx.
Tengine supports different common convolutional neural networks, including SqueezeNet, MobileNet, AlexNet, ResNet, etc. and the optimization strategies such as layered fusion and 8-bit quantization. In addition, it fully develops the performance of Arm CPU by calling the HCL database targeting the optimization of different CPU architectures. For example, it takes merely 111ms to single-threadedly run the MobileNet at mobile devices one time at the RK3399 platform Cortex-A72.
In terms of the innovation and development in respect of AI devices such as IoT devices, intelligent interactive devices, personal computers and robots, Rockchip has showed the leading advantages in technology and ranks among the top 20 in the global AI chip list. In addition to RK3399, the flagship AI chip RK3399Pro attracts equal interests and Rockchip uses the hardware structure design of CPU+GPU+NPU for the first time. The chip integrated NPU (neural network processor) integrates the key technologies of machine vision, speech processing and deep learning, and the on-chip NPU algorithm performance reaches 3.0TOPs, possessing the advantages of high performance, low power consumption and easy development. The RK3399Pro-based EAIDK-610Pro development board has been in full swing. It is a brand new embedded AI application development platform and will provide more comprehensive support for AI developers, the multiple scenarios of partners’ products, the development of whole platforms and ecological layout.
In an era of great reforms to AI chips, the launch of “AIIA DNN benchmark V0.5”, the AIIA’s authoritative test platform, and the announcement of the first-round data will accelerate the technical transformation and evolution of AI chips.