Firefly RM1828SA0. The RM1828SA0 is an AI accelerator module of the inference card type, built around the Rockchip RK1828 SoC and intended for model deployment on embedded devices. Main characteristics — The module integrates a triple-core 64-bit RISC-V processor with FPU, clocked up to 850 MHz, and an NPU offering 20 TOPS in INT8 with support for mixed-precision computation in INT4, INT8, INT16, FP8, FP16 and BF16. It embeds 5GB of on-chip DRAM.
Image processing. The JPEG codec supports encoding and decoding in baseline mode (sequential DCT, JFIF 1.02), and a built-in 2D graphics engine (RGA) performs BT.601/BT.709 pixel format conversion and dithering, with a maximum resolution of 8192×8192 for the source image and 4096×4096 for the target image.
Software and models. The module runs Linux and supports YOLO object-detection algorithms, private deployment of large Transformer-based models such as Gemma, ChatGLM, Qwen and Phi, as well as CNN, RNN and LSTM architectures. RKNN model import and export are supported, with compatibility for TensorFlow, TensorFlow Lite, PyTorch, Keras and PaddlePaddle, plus custom operator development and Docker containerized management.
Interfaces and connectivity. The module uses a 260-pin SODIMM gold-finger interface at 0.5 mm pitch, with an M.2 M-KEY interface. It provides one USB 3.0 DRD port multiplexed with PCIe 2.1 (PCIe by default), one USB 2.0 DRD port, one PCIe 2.1 port and a 4-pin 1.25 mm FAN connector.
Power and environment. The module is powered from 5V DC and 12V DC. It operates between -20°C and 60°C, stores between -20°C and 70°C, with storage humidity of 10% to 90% RH (non-condensing).
Integration. Its weight is 21 g without heat sink and 75 g with heat sink. The module suits integration into robots and embedded AI equipment, and is available on quotation from Atlas Electronics.