Abstract:Multi-head self-attention (attention mechanism) has been employed in a variety of fields such as machine translation, language modeling, and image processing due to its superiority in feature extraction and sequential data analysis. This is benefited from a large number of parameters and sophisticated model architecture behind the attention mechanism. To efficiently deploy attention mechanism on resource-constrained devices, existing works propose to reduce the model size by building a customized smaller model or compressing a big standard model. A customized smaller model is usually optimized for the specific task and needs effort in model parameters exploration. Model compression reduces model size without hurting the model architecture robustness, which can be efficiently applied to different tasks. The compressed weights in the model are usually regularly shaped (e.g. rectangle) but the dimension sizes vary (e.g. differs in rectangle height and width). Such compressed attention mechanism can be efficiently deployed on CPU/GPU platforms as their memory and computing resources can be flexibly assigned with demand. However, for Field Programmable Gate Arrays (FPGAs), the data buffer allocation and computing kernel are fixed at run time to achieve maximum energy efficiency. After compression, weights are much smaller and different in size, which leads to inefficient utilization of FPGA on-chip buffer. Moreover, the different weight heights and widths may lead to inefficient FPGA computing kernel execution. Due to the large number of weights in the attention mechanism, building a unique buffer and computing kernel for each compressed weight on FPGA is not feasible. In this work, we jointly consider the compression impact on buffer allocation and the required computing kernel during the attention mechanism compressing. A novel structural pruning method with memory footprint awareness is proposed and the associated accelerator on FPGA is designed. The experimental results show that our work can compress Transformer (an attention mechanism based model) by 95x. The developed accelerator can fully utilize the FPGA resource, processing the sparse attention mechanism with the run-time throughput performance of 1.87 Tops in ZCU102 FPGA.

Calabash: Accelerating Attention Using a Systolic Array Chain on FPGAs

MAS-Attention: Memory-Aware Stream Processing for Attention Acceleration on Resource-Constrained Edge Devices

Algorithm-hardware Co-design of Attention Mechanism on FPGA Devices

High Performance Scalable FPGA Accelerator for Deep Neural Networks

Automated Systolic Array Architecture Synthesis for High Throughput CNN Inference on FPGAs.

Full-stack Optimization for Accelerating CNNs with FPGA Validation

Integrating FPGA-based hardware acceleration with relational databases

FPGA-Based Near-Memory Acceleration of Modern Data-Intensive Applications

Accelerating Graph Analytics by Co-Optimizing Storage and Access on an FPGA-HMC Platform

A Ubiquitous Machine Learning Accelerator With Automatic Parallelization on FPGA

A Data-Center FPGA Acceleration Platform for Convolutional Neural Networks

An Efficient Hardware Accelerator for Structured Sparse Convolutional Neural Networks on FPGAs

A Hardware Accelerator for Standard Convolution and Depthwise Convolution

Enhancing Long Sequence Input Processing in FPGA-Based Transformer Accelerators Through Attention Fusion

High-Performance FPGA-Based CNN Accelerator with Block-Floating-Point Arithmetic.

Automatic Nested Loop Acceleration on FPGAs Using Soft CGRA Overlay

MALOC: A Fully Pipelined FPGA Accelerator for Convolutional Neural Networks with All Layers Mapped on Chip

A high throughput acceleration for hybrid neural networks with efficient resource management on FPGA

FAMOUS: Flexible Accelerator for the Attention Mechanism of Transformer on UltraScale+ FPGAs

Frequency Improvement of Systolic Array-Based CNNs on FPGAs

Toward Full-Stack Acceleration of Deep Convolutional Neural Networks on FPGAs