Reproduce “A Reconfigurable Framework for AI-FPGA Agent Integration and Acceleration” on OpenResearch

About this paper

Aybars Yunusoglu, Talha Coskun, Hiruna Vishwamith, Murat Isik, I. Can Dikmen

AI-FPGA Agent uses a runtime Q-learning scheduler to assign layers to a CPU or FPGA and coordinate data transfers. On a ResNet-like image-classification CNN tested on 10,000 images, it reports 3.5 ms per image, versus 40.2 ms for a single-threaded CPU and 6.1 ms for a GPU. The comparison measures the combined system; no agent ablation establishes the scheduler’s contribution.

Key insights

• Dynamic, agent-based scheduling with reinforcement learning effectively manages AI workload partitioning on FPGAs, abstracting hardware complexities and adapting to real-time system states.
• Custom FPGA accelerators, optimized with 8-bit quantized arithmetic and pipelined dataflow, significantly improve latency, throughput, and energy efficiency for DNN inference.
• A tightly integrated hardware-software co-design framework can lower the barrier to entry for AI practitioners, making FPGA acceleration more accessible for diverse applications.

Summary by alphaXiv.