傳統的邊緣計算伺服器有固定的計算效能和硬體資源限制，當AIoT的邊緣感測器數量增加，其時間延遲會隨著AI服務需求而上升，一旦超過臨界將使系統運作產生困難。針對低功耗、彈性、可擴充AIoT應用需求，我們設計了一個多主機的AIoT 邊緣計算閘道器，可提供AIoT系統在邊緣端提供彈性化AI計算服務。此一閘道器由一個主控機結合N個可彈性擴充的AI加速從主機，組成成一個彈性架構的N+1的多主機系統。主控機具有快速開發以及佈署的能力，同時藉由負載平衡機制，將邊緣感測器資料串流分散至N個結合AI加速的從主機處理。我們以乳牛牧場管理AIoT應用作為主機的AIoT 邊緣計算閘道器系統的驗證。我們以乳牛偵測、乳牛身分識別和乳牛發情行為辨識等三個AI服務來進行比較實驗。在一段固定時間內，兩個系統同樣運行三個AI服務程式，我們的系統的平均功耗僅為邊緣計算伺服器的21%。在彈性的比較實驗，多主機的AIoT 邊緣計算閘道器在計算需求不斷增加的情況下，藉由彈性增加從主機數量，仍可保持可靠的AI邊緣計算服務效能。實驗結果證明了使用低成本嵌入式平台所建置的多主機的AIoT 邊緣計算閘道器，具有耗電量低的優勢，並且還具備彈性擴充的優勢，更能適應經常變化AI服務的AIoT應用。;Conventional edge computing servers have fixed computing capabilities and limited hardware resources. Increasing numbers of AIoT edge devices will cause longer time delays with a greater demand for AI services. We have designed a flexible and scalable multi-host AIoT edge computing gateway with low power consumption that is capable of providing flexible AI computing services within the edge. The gateway consists of a master host and multiple (N) AI client hosts that can be scaled per demand – creating an N+1 multi-host system with a flexible framework. The master host is equipped with rapid development and deployment capabilities as well as load balancers and is able to distribute the data stream from edge devices to multiple client hosts with AI accelerators. We then tested our multi-host AIoT edge computing gateway by applying it to a comparison experiment of three AI services in dairy farm management AIoT systems: dairy cow detection, identification, and heat detection. The systems ran the three AI services for the same length of time, and the average power consumption of our system was only 21% of that of edge computing servers. For the scalability comparison experiment, the multi-host AIoT gateway was able to still ensure reliable AI edge computing services by increasing the number of client hosts accordingly when computing demands increased. The results confirm that the multi-host edge computing gateway built on low-cost embedded platforms has the advantage as it can maintain flexibility and scalability whilst consuming less power and can adapt to AIoT applications with constantly changing AI services.
