Validation of the NVDLA Architecture via AWS-Based FPGA Co-Simulation Using the AlexNet Model

Bejar D.F.; Jara M.R.; Castillo E.C.V. · 2026 IEEE 17th Latin American Symposium on Circuits and Systems, LASCAS 2026 - Proceedings (2026) · DOI: 10.1109/lascas67804.2026.11457113

Resumen

Deep neural network (DNN) inference presents growing challenges in computational complexity, memory bandwidth, and energy efficiency. To address these demands, this work validates the NVIDIA Deep Learning Accelerator (NVDLA) using a hybrid co-simulation platform deployed on Amazon Web Services (AWS) EC2 F1 FPGA instances. The platform integrates the nv_small RT-level configuration of the NVDLA on the FPGA with a virtual prototype (VP) comprising a QEMU-emulated Arm CPU, SystemC/TLM-based communication, and full Linux runtime support. Validation was conducted through three stages: (1) hardware regression tests to ensure block-level correctness of the NVDLA, (2) software sanity checks to verify system-level integration within the VP, and (3) end-to-end inference of the AlexNet convolutional neural network. All tests passed successfully, with AlexNet inference completing in approximately 112 minutes. These results confirm the functional readiness of the hybrid NVDLA platform and underscore the advantages of VP-FPGA co-simulation for accelerating hardware-software co-design in deep learning accelerator development.

Instituciones: Pontificia Universidad Católica del Perú