Abstract

Energy efficiency is a critical requirement for power-constrained inference workloads, especially in edge scenarios where data are often processed in a frame-oriented manner. While MRAM has been widely explored for low-standby-power systems, its system-level energy behavior under practical inference-driven memory access patterns still requires careful evaluation. A comparative energy analysis is presented for MRAM and DRAM memory access for frame-based MobileNetV3 inference under high-resolution (4K-class) input scenarios, where the memory access intensity is scaled to reflect high-resolution frame workloads rather than pixel-level convolutional dataflows. A counter-based, event-driven energy estimation framework is used to account for DRAM background activity, read/write traffic, and refresh overhead, as well as MRAM read/write energy using a charged-cycle abstraction. The evaluation explicitly incorporates frame gaps, enabling power-gating opportunities and a fair comparison between volatile and non-volatile memory domains under the same workload timeline. Experimental results show that DRAM energy is dominated by continuous background power and refresh overhead, whereas MRAM achieves substantially lower energy per frame due to low standby leakage and effective power gating during frame intervals. These findings highlight the importance of workload-aware memory energy evaluation and suggest that MRAM is a promising memory option for energy-efficient frame-based inference in power-limited systems.

References

• A. Howard et al., “Searching for MobileNetV3,” in Proc. IEEE/CVF Int. Conf. on Computer Vision (ICCV), 2019, pp. 1314–1324, doi: 10.1109/ICCV.2019.00140.
• Arm, “AMBA Specifications (AXI/ACE, AXI4-Stream, APB, etc.),” Arm Developer. [Online]. Available: https://developer.arm.com/architectures/system-architectures/amba/amba-specifications. [Accessed: 13-Feb-2026].
• JEDEC Solid State Technology Association, “JESD79-4D: DDR4 SDRAM Standard,” JEDEC Standard. https://store.accuristech.com/standards/jedec-jesd79-4d. [Accessed: 13-Feb-2026].
• C.-P. Lin, “Designing an AMBA-Compatible MRAM AXI Slave Controller for Modern SoCs,” Medium, Nov. 19, 2025. https://medium.com/@ace.lin0121/designing-an-amba-compatible-mram-axi-slave-controller-for-modern-socs-3cade20bce41. [Accessed: 13-Feb-2026].
• D. Apalkov et al., “Spin-transfer torque magnetic random access memory (STT-MRAM),” ACM J. Emerg. Technol. Comput. Syst. (JETC), vol. 9, no. 2, Art. 13, pp. 1–35, 2013, doi: 10.1145/2463585.2463589.
• M. Sandler et al., “MobileNetV2: Inverted Residuals and Linear Bottlenecks,” in Proc. IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2018, pp. 4510–4520.
• B. Jacob et al., “Quantization and Training of Neural Networks for Efficient Integer-Arithmetic-Only Inference,” in Proc. IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2018.
• Micron Technology, “TN-40-07: Calculating Memory Power for DDR4 SDRAM,” Technical Note. [Online]. https://www.mouser.com/pdfdocs/tn4007_ddr4_power_calculation.pdf. [Accessed: 13-Feb-2026].
• R. Balasubramonian et al., “CACTI 7: New Tools for Interconnect Exploration in Innovative Off-Chip Memories,” ACM Trans. Archit. Code Optim. (TACO), vol. 14, no. 2, Art. 14, 2017, doi: 10.1145/3085572.
• S. Yuasa and D. D. Djayaprawira, “Giant tunneling magnetoresistance in magnetic tunnel junctions with a crystalline barrier,” J. Phys. D: Appl. Phys., vol. 40, p. R337, 2007.