EUVIMEDEUVIMED entdecken
Crossref

A 28-nm Heterogeneous Processor for Programmable Sensor-Fusion Perception

Rakshit Pathak, Akash Pattnaik, Murat Ozatay, Hongyang Jia, Naveen Verma

IEEE Journal of Solid-State Circuits · 2026-11 · 61 (11) · 6706-6719

DOI: 10.1109/jssc.2026.3725199 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Journal of Solid-State Circuits, 2026-11. Verfasst von Rakshit Pathak und 4 weiteren Autor:innen. Die Arbeit nennt 31 Literaturangaben. Gefördert von: Defense Advanced Research Projects Agency (DARPA) Optimum Processing Technology Inside Memory Arrays (OPTIMA) Program, Taiwan Semiconductor Manufacturing Company (TSMC) University Program.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

This article presents a 28-nm heterogeneous programmable perception processor, based on a generalized deep-learning sensor-fusion (DL-SF) model, for combining vision and embedded sensing. The DL-SF model applies across a range of embedded-sensor modalities (acoustic, inertial, temperature, vibration), and its computational characteristics drive processor specialization to achieve simultaneously high programmability and efficiency. The architecture includes: CPU, with custom extensions for programmable DSP-based embedded-sensor feature extraction; configurable convolution accelerator, with parallelism and precision optimized for vision-processing efficiency and throughput requirements; and dense/sparse memory referencing for efficient activation storage. Four bookend applications are demonstrated, with energy reduction relative to DSP application processor: 1) human-activity recognition (classification), achieving baseline iso-accuracy of 98.2% with $10\times $ energy reduction; 2) depth-map generation (regression), achieving baseline iso-error of 0.403 RMSE, with $20\times $ lower energy; 3) human pose estimation (estimation), achieving mean estimation error of 18 mm, with $16\times $ lower energy; and 4) human trajectory forecasting (prediction), achieving average displacement error (ADE) of 26 and final displacement error (FDE) of 46.9, with $28\times $ lower energy.

Förderung

  • Defense Advanced Research Projects Agency (DARPA) Optimum Processing Technology Inside Memory Arrays (OPTIMA) Program
  • Taiwan Semiconductor Manufacturing Company (TSMC) University Program

Zitieren

APA

Pathak, R., Pattnaik, A., Ozatay, M., Jia, H., & Verma, N. (2026). A 28-nm Heterogeneous Processor for Programmable Sensor-Fusion Perception. IEEE Journal of Solid-State Circuits, 61(11), 6706-6719. https://doi.org/10.1109/jssc.2026.3725199

Vancouver

Pathak R, Pattnaik A, Ozatay M, Jia H, Verma N. A 28-nm Heterogeneous Processor for Programmable Sensor-Fusion Perception. IEEE Journal of Solid-State Circuits. 2026;61(11):6706-6719. doi:10.1109/jssc.2026.3725199

BibTeX
@article{Pathak2026,
  author = {Pathak, Rakshit and Pattnaik, Akash and Ozatay, Murat and Jia, Hongyang and Verma, Naveen},
  title = {A 28-nm Heterogeneous Processor for Programmable Sensor-Fusion Perception},
  journal = {IEEE Journal of Solid-State Circuits},
  year = {2026},
  volume = {61},
  number = {11},
  pages = {6706--6719},
  doi = {10.1109/jssc.2026.3725199}
}
Kostenlos starten →