Thursday 04 September 2025
A new framework for generating custom RISC-V extensions has been developed, which could lead to more efficient deployment of artificial intelligence (AI) models on resource-constrained devices such as those found in internet of things (IoT) devices.
The team behind MARVEL, an automated end-to-end framework, has designed a system that can generate model-class aware custom RISC-V extensions tailored specifically for deep neural networks (DNNs). This is significant because traditional AI deployment methods often require hardware modifications or extensive software dependencies, which are not feasible in many IoT devices.
MARVEL’s approach begins by profiling high-level DNN representations in Python and generating an ISA-extended RISC-V core with associated compiler tools. The framework leverages Apache TVM to translate high-level models into optimized C code, Synopsys ASIP Designer for identifying compute-intensive kernels and generating custom extensions, and Xilinx Vivado for FPGA implementation.
The team evaluated MARVEL on popular DNN models such as LeNet-5*, MobileNetV1, ResNet50, VGG16, MobileNetV2, and DenseNet121 using the Synopsys trv32p3 RISC-V core as a baseline. The results show a 2x speedup in inference and up to 2x reduction in energy per inference at a 28.23% area overhead when implemented on an AMD Zynq UltraScale+ ZCU104 FPGA platform.
This development has significant implications for the deployment of AI models on IoT devices, which often have limited processing power and memory. MARVEL’s ability to generate custom extensions tailored specifically for DNNs could enable seamless execution in these environments, without the need for hardware modifications or extensive software dependencies.
The potential applications of MARVEL are vast, from smart home devices to industrial automation systems. By enabling efficient deployment of AI models on resource-constrained devices, MARVEL has the potential to transform various industries and revolutionize the way we interact with our surroundings.
In the future, MARVEL could be used to develop more sophisticated AI models that can be deployed on a wider range of devices, further expanding its impact on various sectors.
Cite this article: “MARVEL: A Framework for Efficient Deployment of AI Models on Resource-Constrained Devices”, The Science Archive, 2025.
Risc-V, Ai, Iot, Marvel, Dnns, Custom Extensions, Fpga, Apache Tvm, Synopsys Asip Designer, Xilinx Vivado
