EAVLA: A Precision- and Sparsity-Adaptive Energy-Aware VLSI Architecture for Artificial-Intelligence-Enabled Embedded Systems
Year: 2025
Type: article
Abstract: Artificial-intelligence inference is increasingly executed directly on embedded platforms to reduce communication latency, preserve privacy, and support always-on sensing. However, the energy cost of multiply-accumulate operations, memory movement, and worst-case fixed-voltage hardware limits the deployment of modern neural networks in battery-powered systems. This paper proposes EAVLA, an energy-aware VLSI accelerator that combines row-stationary local data reuse, layer-wise 4/8/16-bit precision adaptation, zero-value skipping, bank-level clock and power gating, and deadline-aware dynamic voltage and frequency scaling (DVFS). The architecture contains a 16 x 16 array of fusible 4-bit MAC micro-tiles, a 512-KB banked on-chip SRAM hierarchy, a sparse operand decoder, and a lightweight energy manager that selects operating points from layer metadata and application deadlines. A transparent architecture-level analytical model, calibrated to energy trends reported in pre-2022 accelerator literature, is used to compare EAVLA against an area-matched fixed 8-bit dense baseline. Across MobileNetV1, MobileNetV2, SqueezeNet, and a compact DS-CNN keyword-spotting workload, the modeled design reduces energy per inference by 54.3-61.8% (57.4% average) while reducing modeled latency by 23.5-28.9%. The average energy-delay product improves by approximately 2.9x, with an estimated 8.7% logic-area overhead excluding SRAM. Ablation analysis shows that memory locality, sparsity exploitation, mixed precision, and voltage adaptation provide complementary savings. The study demonstrates a practical VLSI co-design direction for embedded AI in which the accelerator explicitly treats energy as a run-time resource rather than a fixed design consequence.
Author: Ratan Babu Telusoori,
Language: English
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7216926595
