MACE is a deep learning inference framework optimized for mobile heterogeneous computing platforms.
MACE addresses the challenge of running neural network models efficiently on resource-constrained mobile devices across Android, iOS, Linux, and Windows. The framework tackles multiple competing demands simultaneously: it optimizes runtime performance through NEON, OpenCL, and Hexagon acceleration along with Winograd convolution algorithms; reduces power consumption via chip-specific scheduling like big.LITTLE and GPU hints; maintains UI responsiveness by breaking OpenCL kernels into preemptible units; minimizes memory usage and library footprint through graph-level allocation optimization and buffer reuse; and protects models through conversion to C++ code and obfuscation techniques. The tool supports multiple input formats including TensorFlow, Caffe, and ONNX models.
MACE suits teams building production mobile AI applications where performance, power efficiency, and model security are critical concerns. It targets recent ARM-based chips from Qualcomm, MediaTek, and Pinecone, making it particularly relevant for Android and iOS deployments. The framework is most appropriate for projects where inference latency and battery impact directly affect user experience, and where protecting proprietary model architectures matters. Organizations should evaluate it against alternatives by consulting the MobileAIBench project, which provides comparative benchmarks across frameworks.
The project maintains active communication channels through GitHub issues, Slack, and QQ groups for user support and feature discussion. Development follows a contribution model that encourages opening issues for discussion before code submissions. The team publishes daily benchmark results against common neural networks on real mobile phones through the MACE Model Zoo, providing concrete performance data for evaluation.
