NAS-Driven Hardware Accelerator Exploration for Edg... | AI Research

NAS-Driven Hardware Accelerator Exploration for Edge AI and Quantization Effects on the Pareto Space addresses the challenge of deploying neural networks on resource-constrained edge devices. The authors propose a three-stage pipeline that automates the design of neural architectures and their subsequent mapping onto reconfigurable hardware, while specifically analyzing how INT4 Post-Training Quantization (PTQ) alters the performance trade-offs of these models.

A Three-Stage Pipeline for Edge Deployment

The authors introduce a workflow that decouples neural architecture design from hardware accelerator optimization:

1. Hardware-Agnostic Frontend: Using the NAS-Bench-201 search space, this stage identifies high-performing architectures based on accuracy and FLOPs using a Pareto rank surrogate. 2. Quantization Bridge: This middleware applies INT4 PTQ to the candidate architectures. It includes a filtering mechanism to remove models that fail to meet performance requirements after quantization, with a feedback loop to trigger a new search if necessary. 3. Hardware-Aware Backend: This stage uses the CGRA4ML tool to map the surviving quantized models onto a Coarse Grain Reconfigurable Array (CGRA). An evolutionary algorithm performs Domain Space Exploration (DSE) to find the optimal hardware configuration—such as processing element array size and memory depth—to minimize latency and area.

Quantization and Architectural Robustness

The researchers conducted an empirical study on all 15,625 architectures in NAS-Bench-201 to understand how INT4 PTQ affects model performance. They observed that models with lower computational complexity (7–47 MFLOPs) are significantly more sensitive to quantization, often showing higher accuracy degradation.
As model complexity increases, this sensitivity decreases. The authors attribute this to a shift in architectural composition: smaller models contain a more heterogeneous mix of operations, whereas larger models increasingly rely on 3x3 convolutions. These convolutions act as a natural low-pass filter, which helps the models remain robust against the noise introduced by reduced-precision arithmetic.

Surrogate Performance and Pareto Stability

A key finding of the study is that the Pareto structure—the set of optimal trade-offs between accuracy and efficiency—is largely preserved when moving from full-precision (FP32) to INT4 quantization.
The authors compared the effectiveness of different surrogate models for guiding the search. They found that an FP32 zero-shot surrogate, which requires no retraining, outperformed a dedicated INT4-trained surrogate in covering the Pareto space. Specifically, the FP32 surrogate achieved a 12.26% improvement in normalized hypervolume for Random Search and a 6.77% improvement for the Multi-Objective Evolutionary Algorithm. This suggests that the ranking landscape of neural architectures remains stable enough under quantization that a surrogate trained on FP32 data is sufficient for effective search.

Franklin Analysis

The research provides a practical framework for edge AI by demonstrating that developers do not necessarily need to incorporate quantization directly into the NAS loop, which is a computationally expensive process. By showing that the FP32 Pareto structure is highly correlated with the quantized one, the authors provide evidence that simpler, decoupled strategies can be effective. The reliance on the CGRA4ML analytical oracle for hardware mapping also allows for rapid design exploration without the need for time-consuming RTL synthesis for every candidate architecture.