11:30 〜 11:45
[PCG20-09] Study on Onboard High-Speed Machine Learning Inference using Dynamically Reconfigurable Processor
キーワード：動的再構成プロセッサ、機械学習、エッジコンピューティング、プラズマ波動
Automatic event classification using a machine learning model is an effective approach for improving the intelligent processing on spacecraft. In general, GPUs (Graphical Processing Units) are used for machine learning and inference on the ground. However, when we implement these functions in a spacecraft, alternative devices are required due to power consumption and thermal design constraints. In this study, we use a dynamically reconfigurable processor (DRP) to achieve small resource and high-speed machine learning inference, aiming for onboard event classification in space.
The Renesas Electronics RZ/V2L microcomputer has a hardware logic component named “DRP-AI” for edge-AI computing. DRP-AI has both features; high-speed processing as FPGA and flexible computing as CPU, through dynamic hardware configuration changes.
In this presentation, we develop a convolutional neural network (CNN) model consisting of six convolutional layers and two fully connected layers, and we evaluate the performance of machine learning inference using DRP-AI. As a result, we confirmed that the inference time using DRP-AI on the RZ/V2L was approximately 20.3 times faster than that using the conventional CPU operating at a clock frequency of 200 MHz. FPGA had an advantage in terms of inference speed, as it flattens computations and performs concurrent processing. However, the DRP-AI had an advantage in terms of resource consumption, as it dynamically changes the hardware configuration. Therefore, we conclude that the RZ/V2L can implement more complex models then the FPGA.
The Renesas Electronics RZ/V2L microcomputer has a hardware logic component named “DRP-AI” for edge-AI computing. DRP-AI has both features; high-speed processing as FPGA and flexible computing as CPU, through dynamic hardware configuration changes.
In this presentation, we develop a convolutional neural network (CNN) model consisting of six convolutional layers and two fully connected layers, and we evaluate the performance of machine learning inference using DRP-AI. As a result, we confirmed that the inference time using DRP-AI on the RZ/V2L was approximately 20.3 times faster than that using the conventional CPU operating at a clock frequency of 200 MHz. FPGA had an advantage in terms of inference speed, as it flattens computations and performs concurrent processing. However, the DRP-AI had an advantage in terms of resource consumption, as it dynamically changes the hardware configuration. Therefore, we conclude that the RZ/V2L can implement more complex models then the FPGA.