Zeroth-order (ZO) optimization is an emerging deep neural network (DNN) training paradigm that offers computational simplicity and memory savings. However, this seemingly promising approach faces a significant and long-ignored challenge. ZO requires generating a substantial number of Gaussian random numbers, which poses significant difficulties and even makes it infeasible for hardware platforms, such as FPGAs and ASICs. In this paper, we identify this critical issue, which arises from the mismatch between algorithm and hardware designers. To address this issue, we proposed PeZO, a perturbation-efficient ZO framework. Specifically, we design random number reuse strategies to significantly reduce the demand for random number generation and introduce a hardware-friendly adaptive scaling method to replace the costly Gaussian distribution with a uniform distribution. Our experiments show that PeZO reduces the required LUTs and FFs for random number generation by 48.6% and 12.7%, and saves at maximum 86% power consumption, all without compromising training performance, making ZO optimization feasible for on-device training. To the best of our knowledge, we are the first to explore the potential of on-device ZO optimization, providing valuable insights for future research.

Superconducting circuits, like Adiabatic Quantum- Flux-Parametron (AQFP), offer exceptional energy efficiency but face challenges in physical design due to sophisticated spacing and timing constraints. Current design tools often neglect the importance of constraint adherence throughout the entire design flow. In this paper, we propose SuperFlow, a fully-customized RTL-to-GDS design flow tailored for AQFP devices. SuperFlow leverages a synthesis tool based on CMOS technology to transform any input RTL netlist to an AQFP-based netlist. Subsequently, we devise a novel place-and-route procedure that simultaneously con-siders wirelength, timing, and routability for AQFP circuits. The process culminates in the generation of the AQFP circuit layout, followed by a Design Rule Check (DR C) to identify and rectify any layout violations. Our experimental results demonstrate that SuperFlow achieves 12.8% wirelength improvement on average and 12.1 % better timing quality compared with previous state- of-the-art placers for AQFP circuits.

This paper introduces SDA, the first effort to adapt the expensive stable diffusion (SD) model for edge FPGA deployment. First, we apply quantization-aware training to quantize its weights to 4 -bit and activations to 8 -bit ($W 4 A 8$) with a negligible accuracy loss. Based on that, we propose a high-performance hybrid systolic array (hybridSA) architecture that natively executes convolution and attention operators across varying quantization bit-widths (e.g., $W 4 A 8$ and all 8 -bit $Q K^{T} V$ in attention). To improve computational efficiency, hybridSA integrates diverse DSP packing techniques into hybrid weightstationary and output-stationary dataflows that are optimized for convolution and attention. It also supports flexible dataflow transitions to address the distinct demands of its output sequence by subsequent nonlinear operators. Moreover, we observe that nonlinear operators become the new performance bottleneck after the acceleration of convolution and attention, and offload them onto the FPGA as well. To reduce the latency of each nonlinear operator, we pipeline its own execution at a fine granularity. To minimize the resource utilization of nonlinear operators, we carefully balance their execution with hybridSA in a coarse-grained pipeline. Experimental results demonstrate that our low-bit ($W 4 A 8$) SDA accelerator on the embedded AMDXilinx ZCU102 FPGA achieves a speedup of $97.3 \times$ (which takes about $\mathrm{2 . 1}$ minutes for one SD inference), compared to the original SD-v1.5 model on the ARM Cortex-A53 CPU (which takes about 3.5 hours for one SD inference). Our SDA project is open sourced here: https://github.com/Michaela1224/SDA_code.

Stochastic rounding is crucial in the low-bit (e.g., 8-bit) training of deep neural networks (DNNs) to achieve high accuracy. One of the drawbacks of prior studies is that they require a large number of high-precision stochastic rounding units (SRUs) to guarantee low-bit DNN accuracy, which involves considerable hardware overhead. In this paper, we use extremely low-bit SRUs (ESRUs) to save a large number of hardware resources during low-bit DNN training. However, a naively designed ESRU introduces a biased distribution of random numbers, causing accuracy degradation. To address this issue, we further propose an ESRU design with a plateau-shape distribution. The plateau-shape distribution in our ESRU design is implemented with the combination of an LFSR (linear-feedback shift register) and an inverted LFSR, which avoids LFSR packing and turns an inherent LFSR drawback into an advantage in our efficient ESRU design. Experimental results using state-of-the-art DNN models demonstrate that, compared to the prior 24-bit SRU with 24-bit pseudo-random number generators (PRNG), our 8-bit ESRU with 3-bit PRNG reduces the SRU hardware resource usage by 9.75x while achieving slightly higher accuracy.

Shaoyi Huang, Dongkuan Xu, Ian Yen, Yijue Wang, Sung-En Chang, Bingbing Li, Shiyang Chen, Mimi Xie, Sanguthevar Rajasekaran, Hang Liu, Caiwen Ding. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2022.

Stochastic rounding is crucial in the training of low-bit deep neural networks (DNNs) to achieve high accuracy. Unfortunately, prior studies require a large number of high-precision stochastic rounding units (SRUs) to guarantee the low-bit DNN accuracy, which involves considerable hardware overhead. In this paper, we propose an automated framework to explore hardware-efficient low-bit SRUs (ESRUs) that can still generate high-quality random numbers to guarantee the accuracy of low-bit DNN training. Experimental results using state-of-the-art DNN models demonstrate that, compared to the prior 24-bit SRU with 24-bit pseudo random number generator (PRNG), our 8-bit with 3-bit PRNG reduces the SRU resource usage by 9.75× while achieving a higher accuracy.

With the trend to deploy Deep Neural Network (DNN) inference models on edge devices with limited resources, quantization techniques have been widely used to reduce on-chip storage and improve computation throughput. However, existing DNN quantization work deploying quantization below 8-bit may be either suffering from evident accuracy loss or facing a big gap between the theoretical improvement of computation throughput and the practical inference speedup. In this work, we propose a general framework, called FILM-QNN, to quantize and accelerate multiple DNN models across different embedded FPGA devices. First, we propose the novel intra-layer, mixed-precision quantization algorithm that assigns different precisions onto the filters of each layer. The candidate precision levels and assignment granularity are determined from our empirical study with the capability of preserving accuracy and improving hardware parallelism. Second, we apply multiple optimization techniques for the FPGA accelerator architecture in support of quantized computations, including DSP packing, weight reordering, and data packing, to enhance the overall throughput with the available resources. Moreover, a comprehensive resource model is developed to balance the allocation of FPGA computation resources (LUTs and DSPs) as well as data transfer and on-chip storage resources (BRAMs) to accelerate the computations in mixed precisions within each layer. Finally, to improve the portability of FILM-QNN, we implement it using Vivado High-Level Synthesis (HLS) on Xilinx PYNQ-Z2 and ZCU102 FPGA boards. Our experimental results of ResNet-18, ResNet-50, and MobileNet-V2 demonstrate that the implementations with intra-layer, mixed-precision (95% of 4-bit weights and 5% of 8-bit weights, and all 5-bit activations) can achieve comparable accuracy (70.47%, 77.25%, and 65.67% for the three models) as the 8-bit (and 32-bit) versions and comparable throughput (214.8 FPS, 109.1 FPS, and 537.9 FPS on ZCU102) as the 4-bit designs.

Deep Neural Networks (DNNs) have achieved extraordinary performance in various application domains. To support diverse DNN models, efficient implementations of DNN inference on edge-computing platforms, e.g., ASICs, FPGAs, and embedded systems, are extensively investigated. Due to the huge model size and computation amount, model compression is a critical step to deploy DNN models on edge devices. This paper focuses on weight quantization, a hardware-friendly model compression approach that is complementary to weight pruning.Unlike existing methods that use the same quantization scheme for all weights, we propose the first solution that applies different quantization schemes for different rows of the weight matrix. It is motivated by (1) the distribution of the weights in the different rows are not the same; and (2) the potential of achieving better utilization of heterogeneous FPGA hardware resources. To achieve that, we first propose a hardware-friendly quantization scheme named sum-of-power-of-2 (SP2) suitable for Gaussian-like weight distribution, in which the multiplication arithmetic can be replaced with logic shifter and adder, thereby enabling highly efficient implementations with the FPGA LUT resources. In contrast, the existing fixed-point quantization is suitable for Uniform-like weight distribution and can be implemented efficiently by DSP. Then to fully explore the resources, we propose an FPGA-centric mixed scheme quantization (MSQ) with an ensemble of the proposed SP2 and the fixed-point schemes. Combining the two schemes can maintain, or even increase accuracy due to better matching with weight distributions.For the FPGA implementations, we develop a parameterized architecture with heterogeneous Generalized Matrix Multiplication (GEMM) cores-one using LUTs for computations with SP2 quantized weights and the other utilizing DSPs for fixed-point quantized weights. Given the partition ratio among the two schemes based on resource characterization, MSQ quantization training algorithm derives an optimally quantized model for the FPGA implementation. We evaluate our FPGA-centric quantization framework across multiple application domains. With optimal SP2/fixed-point ratios on two FPGA devices, i.e., Zynq XC7Z020 and XC7Z045, we achieve performance improvement of 2.1 × -4.1 × compared to solely exploiting DSPs for all multiplication operations. In addition, the CNN implementations with the proposed MSQ scheme can achieve higher accuracy and comparable hardware utilization efficiency compared to the state-of-the-art designs.

This work proposes a novel Deep Neural Network (DNN) quantization framework, namely RMSMP, with a Row-wise Mixed-Scheme and Multi-Precision approach. Specifically, this is the first effort to assign mixed quantization schemes and multiple precisions within layers – among rows of the DNN weight matrix, for simplified operations in hardware inference, while preserving accuracy. Furthermore, this paper makes a different observation from the prior work that the quantization error does not necessarily exhibit the layer-wise sensitivity, and actually can be mitigated as long as a certain portion of the weights in every layer are in higher precisions. This observation enables layer-wise uniformality in the hardware implementation towards guaranteed inference acceleration, while still enjoying row-wise flexibility of mixed schemes and multiple precisions to boost accuracy. The candidates of schemes and precisions are derived practically and effectively with a highly hardware-informative strategy to reduce the problem search space.With the offline determined ratio of different quantization schemes and precisions for all the layers, the RMSMP quantization algorithm uses Hessian and variance based method to effectively assign schemes and precisions for each row. The proposed RMSMP is tested for the image classification and natural language processing (BERT) applications, and achieves the best accuracy performance among state-of-the-arts under the same equivalent precisions. The RMSMP is implemented on FPGA devices, achieving 3.65× speedup in the end-to-end inference time for ResNet-18 on ImageNet, comparing with the 4-bit Fixed-point baseline.

We consider a generalization of mixed regression where the response is an additive combination of several mixture components. Standard mixed regression is a special case where each response is generated from exactly one component. Typical approaches to the mixture regression problem employ local search methods such as Expectation Maximization (EM) that are prone to spurious local optima. On the other hand, a number of recent theoretically-motivated \emph{Tensor-based methods} either have high sample complexity, or require the knowledge of the input distribution, which is not available in most of practical situations. In this work, we study a novel convex estimator \emph{MixLasso} for the estimation of generalized mixed regression, based on an atomic norm specifically constructed to regularize the number of mixture components. Our algorithm gives a risk bound that trades off between prediction accuracy and model sparsity without imposing stringent assumptions on the input/output distribution, and can be easily adapted to the case of non-linear functions. In our numerical experiments on mixtures of linear as well as nonlinear regressions, the proposed method yields high-quality solutions in a wider range of settings than existing approaches.

The latent feature model (LFM), proposed in (Griffiths \& Ghahramani, 2005), but possibly with earlier origins, is a generalization of a mixture model, where each instance is generated not from a single latent class but from a combination of latent features. Thus, each instance has an associated latent binary feature incidence vector indicating the presence or absence of a feature. Due to its combinatorial nature, inference of LFMs is considerably intractable, and accordingly, most of the attention has focused on nonparametric LFMs, with priors such as the Indian Buffet Process (IBP) on infinite binary matrices. Recent efforts to tackle this complexity either still have computational complexity that is exponential, or sample complexity that is high-order polynomial w.r.t. the number of latent features. In this paper, we address this outstanding problem of tractable estimation of LFMs via a novel atomic-norm regularization, which gives an algorithm with polynomial run-time and sample complexity without impractical assumptions on the data distribution.