With the rise of IoT, e-commerce, and 5G, the demands of secure communications and identity authentications dramatically increase, which largely rely on high-volume cryptographic computing. Meanwhile, driven by deep learning, the embedded GPU system is rapidly evolving. In this paper, we discuss the feasibility of turning the lightweight and energy efficient system into a cryptographic co-processor, where security and performance are two daunting challenges. From the aspect of security, we leverage the available resources in the embedded GPU system to achieve on-chip uninterrupted cryptographic computing, secure key storage, and trusted system bootstrapping. From the aspect of performance, targeting the prevailing digital signature algorithm Ed25519, we develop an entire framework to make full use of the system’s cryptographic computing power, including the Ed25519 implementation with embedded GPU acceleration and a high-performance network processing architecture. In Jetson Xavier and Jetson Xavier NX, we implement a prototype called SECCEG and conduct comprehensive experiments to evaluate its performance. SECCEG can serve as a network cryptographic accelerator via TCP/IP stack with ops signature generation and ops for signature verification. At the performance-power ratio, SECCEG achieves 122 kops/W and 35.6 kops/W for signature generation and verification, respectively, which is 1 to 2 multitude higher than ARM CPU, FPGA and discrete GPU implementations.