부산대학교 도서관

With the explosive data volume generated by satellites, in-space computing becomes indispensable to fully exploit the value of such data. Similar to today’s computing paradigms on the ground, a key step to the success of in-space computing is building distributed, micro data centers in space, which consist of many server machines. In this work, we propose a novel server architecture that is comprised of massive, low-power system-on-chips (SoCs). Through quantitative analysis, we confirm that such a server significantly outperforms conventional servers (Intel CPU, NVIDIA GPU, etc) on three critical metrics of in-space computing, i.e., energy efficiency, weight, and volume. Notably, we also demonstrate that the three concerned metrics above can be reduced to only one (energy efficiency) as the accompanied solar panels to provide the needed energy often overwhelm the server itself in weight and volume. We further dive into two specific applications that are representative of in-space computing: video processing and deep learning inference. Through state-of-the-art software and benchmarks, we reveal that our SoC-based server, especially with its heterogeneous processors (GPU, DSP, and MediaCodec), can remarkably improve the energy efficiency compared to conventional servers.