Work

Um framework para avaliação acelerada por FPGA de circuitos aritméticos aproximados em arranjos sistólicos

Year: 2026
Type: dissertation
Abstract: The growing demand for deep learning applications across edge and cloud environments has intensified the adoption of domain-specific architectures, particularly systolic arraybased accelerators for efficient matrix computations. While these architectures achieve high throughput and energy efficiency, their scalability is fundamentally constrained by the cost of multiply–accumulate (MAC) units, which dominate area and power consumption. Approximate computing offers a promising solution by exploiting the intrinsic error resilience of neural networks, enabling more efficient arithmetic units with minimal impact on inference accuracy. But diverse approximate circuits have a variable impact depending on neural network architecture and dataset, thus requiring the automation of exploration to effectively determine approximation impact on a target workload. In this scenario, we propose a framework for cross-layer evaluation of approximate arithmetic in systolic DNN accelerators. The framework integrates hardware synthesis, FPGA-based prototyping, and neural network evaluation to systematically analyze the trade-offs between power, area, and accuracy. By generating Pareto-optimal design points, our framework enables informed selection of approximation strategies tailored to specific workloads and architectures. Our main results include the evaluation of the thirteen circuits of EvoApproxLib-lite with three deep convolutional neural networks.
Language: Portuguese
Open Access status: green
OpenAlex ID: W7212331956
Cites: 27
Cited by:
Related to: 10