eIQ Neutron: Redefining Edge-AI Inference with Integrated NPU and Compiler Innovations

Neural Processing Units (NPUs) are key to enabling efficient AI inference in resource-constrained edge environments. While peak tera operations per second (TOPS) is often used to gauge performance, it poorly reflects real-world performance and typically rather correlates with higher silicon cost. To address this, architects must focus on maximizing compute utilization, without sacrificing flexibility. This paper presents the eIQ Neutron efficient-NPU, integrated into a commercial flagship MPU, alongside co-designed compiler algorithms. The architecture employs a flexible, data-driven design, while the compiler uses a constrained programming approach to optimize compute and data movement based on workload characteristics. Compared to the leading embedded NPU and compiler stack, our solution achieves an average speedup of 1.8x (4x peak) at equal TOPS and memory resources across standard AI-benchmarks. Even against NPUs with double the compute and memory resources, Neutron delivers up to 3.3x higher performance.

Design and Optimization of Residual Neural Network Accelerators for Low-Power FPGAs Using High-Level Synthesis

Residual neural networks are widely used in computer vision tasks. They enable the construction of deeper and more accurate models by mitigating the vanishing gradient problem. Their main innovation is the residual block which allows the output of one layer to bypass one or more intermediate layers and be added to the output of a later layer. Their complex structure and the buffering required by the residual block make them difficult to implement on resource-constrained platforms. We present a novel design flow for implementing deep learning models for field programmable gate arrays optimized for ResNets, using a strategy to reduce their buffering overhead to obtain a resource-efficient implementation of the residual layer. Our high-level synthesis (HLS)-based flow encompasses a thorough set of design principles and optimization strategies, exploiting in novel ways standard techniques such as temporal reuse and loop merging to efficiently map ResNet models, and potentially other skip connection-based NN architectures, into FPGA. The models are quantized to 8-bit integers for both weights and activations, 16-bit for biases, and 32-bit for accumulations. The experimental results are obtained on the CIFAR-10 dataset using ResNet8 and ResNet20 implemented with Xilinx FPGAs using HLS on the Ultra96-V2 and Kria KV260 boards. Compared to the state-of-the-art on the Kria KV260 board, our ResNet20 implementation achieves 2.88X speedup with 0.5% higher accuracy of 91.3%, while ResNet8 accuracy improves by 2.8% to 88.7%. The throughputs of ResNet8 and ResNet20 are 12971 FPS and 3254 FPS on the Ultra96 board, and 30153 FPS and 7601 FPS on the Kria KV26, respectively. They Pareto-dominate state-of-the-art solutions concerning accuracy, throughput, and energy.