HAWX: A Hardware-Aware FrameWork for Fast and Scalable ApproXimation of DNNs

Samira Nazari; Mohammad Saeed Almasi; Mahdi Taheri; Ali Azarpeyvand; Ali Mokhtari; Ali Mahani; Christian Herglotz

cs.LGFeb 18, 2026

HAWX: A Hardware-Aware FrameWork for Fast and Scalable ApproXimation of DNNs

Authors: Samira Nazari, Mohammad Saeed Almasi, Mahdi Taheri, Ali Azarpeyvand, Ali Mokhtari, Ali Mahani, Christian Herglotz

Organizations: University of Zanjan, Zanjan, Iran · Brandenburg Technical University, Cottbus, Germany · Tallinn University of Technology, Tallinn, Estonia · Shahid Bahonar University, Kerman, Iran

Abstract

This work presents HAWX, a hardware-aware scalable exploration framework that employs multi-level sensitivity scoring at different DNN abstraction levels (operator, filter, layer, and model) to guide selective integration of heterogeneous AxC blocks. Supported by predictive models for accuracy, power, and area, HAWX accelerates the evaluation of candidate configurations, achieving over 23* speedup in a layer-level search with two candidate approximate blocks and more than (3106) speedup at the filter-level search only for LeNet-5, while maintaining accuracy comparable to exhaustive search. Experiments across state-of-the-art DNN benchmarks such as VGG-11, ResNet-18, and EfficientNetLite demonstrate that the efficiency benefits of HAWX scale exponentially with network size. The HAWX hardware-aware search algorithm supports both spatial and temporal accelerator architectures, leveraging either off-the-shelf approximate components or customized designs.

Figures & tables

Fig. 1 : A high-level flow of the proposed methodology

Fig. 2 : Activation distribution of LeNet-5 on representative inputs

Fig. 3 : hawx(o) distribution in LeNet-5

Model	AxC Block	Filters		Layers
		1	2	1	2
LeNet-5	KV8	6.46E-06	1.54E-05	1.18E-05	1.60E-05
LeNet-5	L2N	7.87E-06	1.78E-05	1.33E-05	1.71E-05
VGG-11	KV8	9.16E-05	8.53E-05	7.90E-05	1.30E-06
VGG-11	L2N	9.86E-05	9.37E-05	8.50E-05	1.30E-05
ResNet-18	KV8	3.00E-04	1.00E-04	1.81E-04	2.43E-04

TABLE I : Comparison of KV8 and L2N

hawx(f)

and

hawx(l)

across sample filters and layers.

Panel of Figure 5 — Fig. 4 : Comparison of Accuracy (blue) and $hawx(model)$ (red) across DNNs.

Model	Configuration	Acc.	Pwr.	Area
LeNet-5	1-5:KV6	97.80	1.000	1.000
	1-5:KV9	97.77	0.965	0.939
	1-2:(KV9,KVP) \| 3-5:KV9	97.77	0.938	0.921
	1-4:KVP \| 5:L2J	97.54	0.854	0.870
	1-2:(KVP,L2J) \| 3-5:KVP	97.54	0.826	0.850
	1-5:L2N	95.58	0.296	0.390

TABLE II : Normalized trade-off between accuracy (Acc.), power (Pwr.), and area (Area).

Panel of Figure 7 — Fig. 5 : Trade-off between accuracy, power, and area for dataflow and PERUN architectures.

Neural Network Hardware Acceleration

See all 38 papers

AxMoE: Characterizing the Impact of Approximate Multipliers on Mixture-of-Experts DNN Architectures
May 6, 2026Omkar B Shende, Marcello Traiola, Gayathri AnanthanarayananEfficient Neural Network Inference Edge Inference
Dept. of Computer Science and Engineering · Univ Rennes, CNRS, Inria, IRISA
DASH: Fast Differentiable Architecture Search for Hybrid Attention in Minutes on a Single GPU
May 20, 2026Weizhe Chen, Miao Zhang, Junpeng Jiang +3LLM Inference Efficiency Attention Mechanisms
Harbin Institute of Technology (Shenzhen)
FAME: An FPGA-Based Platform for Approximate Multipliers Evaluation with Pattern-Guided DNN Retraining
Sep 15, 2026Rappy Saha, Nima Amirafshar, Jude Haris +2FPGA-Based Neural Network Acceleration Approximate Computing
University of Glasgow, UK · Heidelberg University, Germany

May 6, 2026J/K move · Enter open · S save

AxMoE: Characterizing the Impact of Approximate Multipliers on Mixture-of-Experts DNN Architectures

Omkar B Shende, Marcello Traiola, Gayathri Ananthanarayanan

Dept. of Computer Science and Engineering · Univ Rennes, CNRS, Inria, IRISA

Efficient Neural Network Inference Edge Inference Mixture-of-Experts Models Approximate Computing

Open paper AskPDF arXiv

Deep neural network (DNN) inference at the edge demands simultaneous improvements in accuracy, computational efficiency, and energy consumption. Approximate computing and Mixture-of-Experts (MoE) architectures have each been studied as independent routes towards efficient inference, the former by replacing exact arithmetic with low-power approximate multipliers, the latter by routing inputs through specialized expert sub-networks to enable conditional computation. However, their interaction remains entirely unexplored. This paper presents AxMoE, the first study of the impact of approximate multiplication on MoE DNN architectures. We evaluate three MoE variants: Hard MoE, Soft MoE, and Cluster MoE against dense baselines across three CNN architectures (ResNet-20, VGG11_bn, VGG19_bn) on CIFAR-100 and a Vision Transformer (ViT-Small) on Tiny ImageNet-200 dataset, using eight 8-bit signed multipliers (including one exact baseline) from the EvoApproxLib library. Results show that, without retraining, the Dense baseline is the most resilient topology across all CNN architectures, whereas on ViT-Small, all topologies degrade at comparable rates regardless of routing strategy. After approximate-aware retraining, recovery varies substantially across architectures, topologies, and multipliers. ResNet-20 achieves full recovery across the entire multiplier range, whereas VGG architectures recover at moderate multipliers but fail irreversibly at aggressive ones for all topologies except Cluster MoE on VGG11_bn; on ViT-Small, Hard MoE outperforms Dense under aggressive approximation at equal normalized inference cost. These results pave the way for future approximate MoE hardware-software co-design strategies.

Abstract

Neural Network Hardware Acceleration

Explore similar work

AxMoE: Characterizing the Impact of Approximate Multipliers on Mixture-of-Experts DNN Architectures

DASH: Fast Differentiable Architecture Search for Hybrid Attention in Minutes on a Single GPU

FAME: An FPGA-Based Platform for Approximate Multipliers Evaluation with Pattern-Guided DNN Retraining