Richie Li | Searcharxiv

arXiv2025

Dataflow & Tiling Strategies in Edge-AI FPGA Accelerators: A Comprehensive Literature Review

Edge-AI applications demand high-throughput, low-latency inference on FPGAs under tight resource and power constraints. This survey provides a comprehensive review of two key architectural decisions for FPGA-based neural network accelerators: (i) the dataflow (the order and manner in which data is moved and reused on chip), and (ii) the tiling/blocking strategy (how large tensors are partitioned to fit on-chip). We first present a broadened taxonomy of canonical dataflow styles: Weight-Stationary, Output-Stationary, Row-Stationary, and No-Local-Reuse, including formal definitions, pseudocode/diagrams, and real FPGA examples. We then discuss analytical frameworks (MAESTRO, Timeloop) and compare them with a concise feature table, illustrating how they model reuse, performance, and hardware costs, and include a case study of a 3x3 convolution layer to demonstrate typical tool outputs. Next, we detail multi-level tiling and loop unrolling/pipelining strategies for FPGAs, clarifying how each memory tier (registers, LUTRAM, BRAM, HBM) can be exploited. Our four case studies - FINN, FINN-R, FlightLLM, and SSR - highlight distinct dataflows (from binary streaming to hybrid sparse transformations) and tiling patterns. We include a unified comparison matrix covering platform, precision, throughput, resource utilization, and energy efficiency, plus small block diagrams for each design. We conclude by examining design automation trade-offs among HLS, DSL, and hand-coded RTL, offering a "lessons learned" summary box, and charting future research directions in partial reconfiguration, hybrid dataflows, and domain-specific compiler flows for next-generation edge AI FPGA accelerators.

cs.AR↗

arXiv2025

Design and Implementation of an FPGA-Based Hardware Accelerator for Transformer

Transformer-based large language models (LLMs) rely heavily on intensive matrix multiplications for attention and feed-forward layers, with the Q, K, and V linear projections in the Multi-Head Self-Attention (MHA) module constituting a decisive performance bottleneck. In this work, we introduce a highly optimized tiled matrix multiplication accelerator on a resource-constrained Xilinx KV260 FPGA that not only addresses this challenge but sets a new standard for efficiency and performance. Our design exploits persistent on-chip storage, a robust two-level tiling strategy for maximal data reuse, and a systolic-like unrolled compute engine that together deliver unparalleled speed and energy efficiency. Integrated with DistilBERT for Q, K, and V projections, our accelerator achieves an unequivocal 7x speedup over ARM CPU implementations (PyTorch) and an extraordinary 200x improvement over naive NumPy, reaching a throughput of up to 3.1~GFLOPs for matrix multiplications on (64,768) x (768,3072) matrices while operating at a conservative 100 MHz. These results decisively demonstrate the transformative potential of FPGA-based acceleration for critical Transformer operations, paving the way for scalable and energy-efficient deep learning inference on edge devices.

cs.AR↗