Work

Accelerating Quantized DNNs with Dedicated Hardware Accelerators and RISC-V Processors Using Precision-Scalable Multipliers

Year: 2024
Type: dissertation
Abstract: Mixed-Precision Quantization (MPQ) and Transprecision Computing (TC) represent two valuable techniques used to optimize Deep Neural Networks (DNNs) inference.They aim at minimizing the number of activation and weight bits for each DNN layer during training, and dynamically adjusting the numerical precision during runtime, respectively.Their goal is to find an optimal balance between accuracy, latency, and energy consumption.Implementing MPQ and TC in practice necessitates the use of Precision-Scalable (PS) and reconfigurable hardware.This aspect constitutes the primary topic of this thesis.Given that Deep Learning (DL) algorithms essentially involve scalar multiplications and dot products for executing convolutions and matrix multiplications, our attention is on PS multipliers.Specifically, we focus on two main categories of PS multiplier architectures, Sum-Apart (SA) and Sum-Together (ST), and we integrate them into the Multiplyand-Accumulate (MAC) units of DNN accelerators and low-power extreme-edge RISC-V processors.In these multipliers, N multiplications are computed in parallel in a Single Instruction Multiple Data (SIMD) fashion, with operands on 16/N bits, where N = 1, 2, 4. While SA multipliers keep the results separate from each other, ST multipliers accumulate the results of low-precision multiplication internally, eliminating the need for an external adder.Consequently, they enable support for MPQ and TC and, at the same time, accelerate MAC operations by a factor of up to N compared to conventional full-precision 16-bit multipliers.Our study provides a comprehensive comparison of the main ST multipliers in the literature.We begin with an overview of State-of-the-Art (SoA) ST multiplier architectures.Next, we introduce three new designs: one optimizing the critical path of a Baugh-Wooley (BW) ST multiplier, another derived from High-Level Synthesis (HLS), and the third based the Booth architecture.We evaluate their performance, power, and area (PPA) characteristics across a wide clock frequency range, after normalizing all the architectures to support 16, 8, and 4 bits of precision.The key finding reveals no single winner that satisfies all PPA scenarios, but rather a set of optimal ST multipliers depending on specific PPA constraints.Our research also contributes to the advancement of ST-based DNN hardware accelerators by proposing implementations for 2D-Convolution (2D-Conv), Depth-wise Convolution (DW-Conv), and Fully-Connected (FC) layers.These
Source: PORTO@Iris
Language: English
Open Access status: green
OpenAlex ID: W7217249430
Cites: 95
Cited by: 2
Related to: 10