Reproduce “No Attention, No Problem: DPU-Aware Attention Approximation in Modern YOLO on FPGA” on OpenResearch

About this paper

Suraj Karki, Qazi Arbab Ahmed, Thorsten Jungeblut

A DPU-compatible approximation of YOLO spatial attention enables recent standard and oriented detectors to run on an AMD/Xilinx FPGA without unsupported operators. On a ZCU104, the adapted models reached up to 34.05 FPS for standard detection and 29.55 FPS for oriented detection at B4096, while operating within a measured DPU power range of 4.792–8.760 W, at the cost of several points of accuracy after architectural modification and 8-bit quantization.

Key insights

• Replacing unsupported attention operations with DPU-native primitives preserves the general feature-weighting behavior, but the approximation is not accuracy-neutral: architectural changes reduce performance by roughly three mAP points on average, with additional degradation after quantization.
• Larger DPU configurations improve accelerator-level throughput, yet end-to-end gains are limited by ARM-side preprocessing and especially non-maximum suppression; faster DPU execution therefore does not translate proportionally into application throughput.
• YOLOv26 benefits from a simplified detection design, achieving the highest reported end-to-end speed, while YOLOv11 generally retains accuracy more effectively after FPGA deployment.

Summary by alphaXiv.