An IoT Endpoint System-on-Chip for Secure and Energy-Efficient Near-Sensor Analytics

Francesco Conti, Robert Schilling, Pasquale Davide Schiavone, Antonio Pullini, Davide Rossi, Frank Kagan Gürkaynak, Michael Muehlberghuber, Michael Gautschi, Igor Loi, Germain Haugou, Stefan Mangard, Luca Benini

Abstract

Near-sensor data analytics is a promising direction for IoT endpoints, as it minimizes energy spent on communication and reduces network load - but it also poses security concerns, as valuable data is stored or sent over the network at various stages of the analytics pipeline. Using encryption to protect sensitive data at the boundary of the on-chip analytics engine is a way to address data security issues. To cope with the combined workload of analytics and encryption in a tight power envelope, we propose Fulmine, a System-on-Chip based on a tightly-coupled multi-core cluster augmented with specialized blocks for compute-intensive data processing and encryption functions, supporting software programmability for regular computing tasks. The Fulmine SoC, fabricated in 65nm technology, consumes less than 20mW on average at 0.8V achieving an efficiency of up to 70pJ/B in encryption, 50pJ/px in convolution, or up to 25MIPS/mW in software. As a strong argument for real-life flexible application of our platform, we show experimental results for three secure analytics use cases: secure autonomous aerial surveillance with a state-of-the-art deep CNN consuming 3.16pJ per equivalent RISC op; local CNN-based face detection with secured remote recognition in 5.74pJ/op; and seizure detection with encrypted data collection from EEG within 12.7pJ/op.

Introduction

IoT devices collect massive amounts of rich data, including video, audio, and biosignals. Transmitting this raw data to the cloud for processing is often prohibitive due to battery constraints and network bandwidth limitations. "Near-sensor analytics"—processing data directly on the edge device to extract actionable insights—offers a solution, but it introduces significant security vulnerabilities. Sensitive distilled data and valuable machine learning models (such as CNN weights) stored on-device require robust encryption.

The challenge lies in the computational intensity: running deep learning and cryptographic algorithms simultaneously requires billions of operations per second. Standard microcontrollers lack the necessary throughput, while existing hardware accelerators are often too power-hungry or inflexible to operate within the typical IoT power budget of 20-30 mW.

Architecture and Methods

The Fulmine SoC is a 65nm heterogeneous multi-core system designed to bridge the gap between performance, security, and energy efficiency. Its architecture centers on a tightly-coupled multi-core cluster that minimizes data movement overhead.

Architecture of the Fulmine SoC and its multi-core cluster.

Key architectural innovations include:

  • Tightly-Coupled Shared Memory (TCDM): Instead of traditional bus-based communication, Fulmine utilizes a shared Level 1 TCDM. Four OpenRISC cores and two specialized hardware accelerators share this memory, enabling "zero-copy" data exchange and drastically reducing energy consumption associated with data movement.
  • Hardware Cryptography Engine (HWCRYPT): A dedicated block supporting AES-128 (including XTS mode for secure external memory storage) and Keccak (SHA-3). This is critical because AES-XTS has internal data dependencies that make software parallelization inefficient.
  • Hardware Convolution Engine (HWCE): A specialized block for CNN processing that features precision scalability. It can dynamically reconfigure from 16-bit to 8-bit or 4-bit weights, allowing it to compute multiple convolutions simultaneously to maximize throughput and memory bandwidth utilization.
  • Aggressive Power Management: The chip employs independent voltage and frequency domains, granular clock-gating, and a deep-sleep mode to enable extreme duty-cycling for low-power operation.

Application Workloads

The platform was evaluated using three real-world secure analytics use cases:

  1. Secure Autonomous Aerial Surveillance: Running a ResNet-20 CNN on 224x224 images for object classification.
  2. Local Face Detection: Utilizing 12-net and 24-net CNNs to detect faces locally before encrypting and transmitting the results to a remote server.
  3. Seizure Detection: Processing multi-channel EEG biosignals via Principal Component Analysis (PCA), Wavelet Transforms, and Support Vector Machines (SVM).

Key Results and Performance

Fulmine achieves state-of-the-art efficiency, consuming less than 20 mW on average at 0.8V. In a mixed workload environment involving CNNs, encryption, and software filtering, the system achieves an efficiency of 5.74 pJ per equivalent RISC operation.

Power management scheme and measured power consumption across operating modes.

Compared to existing solutions like SleepWalker, Fulmine executes the same workload 89x faster while maintaining superior energy efficiency. For the UAV surveillance use case, a secure ResNet-20 inference (with on-the-fly encryption/decryption of weights and partial results) consumes only 27 mJ. This allows a nano-drone to run continuous surveillance for a 7-minute flight using less than 0.25% of its total battery capacity. Similarly, the EEG seizure detection pipeline requires only 0.18 mJ per inference, enabling over two years of continuous monitoring on a standard pacemaker battery.

Discussion and Future Work

While Fulmine significantly advances edge analytics, two primary bottlenecks remain. First, in ultra-efficient 4-bit CNN modes, the hardware accelerator can saturate the 4-port memory bandwidth, suggesting that future designs may require wider memory interfaces. Second, energy consumption remains dominated by external memory accesses (Flash/FRAM) for heavy workloads like ResNet-20.

Future research directions include moving toward System-on-Package (SoP) designs. By vertically stacking the sensor, memory, and the SoC in a single package, the energy cost of die-to-die connections can be drastically reduced, further extending the battery life of secure IoT endpoints.

Summary updated

Abstract

Near-sensor data analytics is a promising direction for IoT endpoints, as it minimizes energy spent on communication and reduces network load - but it also poses security concerns, as valuable data is stored or sent over the network at various stages of the analytics pipeline. Using encryption to protect sensitive data at the boundary of the on-chip analytics engine is a way to address data security issues. To cope with the combined workload of analytics and encryption in a tight power envelope, we propose Fulmine, a System-on-Chip based on a tightly-coupled multi-core cluster augmented with specialized blocks for compute-intensive data processing and encryption functions, supporting software programmability for regular computing tasks. The Fulmine SoC, fabricated in 65nm technology, consumes less than 20mW on average at 0.8V achieving an efficiency of up to 70pJ/B in encryption, 50pJ/px in convolution, or up to 25MIPS/mW in software. As a strong argument for real-life flexible application of our platform, we show experimental results for three secure analytics use cases: secure autonomous aerial surveillance with a state-of-the-art deep CNN consuming 3.16pJ per equivalent RISC op; local CNN-based face detection with secured remote recognition in 5.74pJ/op; and seizure detection with encrypted data collection from EEG within 12.7pJ/op.
BibTeX
@article{conti2017iot,
  author = {Francesco Conti and Robert Schilling and Pasquale Davide Schiavone and Antonio Pullini and Davide Rossi and Frank Kagan Gürkaynak and Michael Muehlberghuber and Michael Gautschi and Igor Loi and Germain Haugou and Stefan Mangard and Luca Benini},
  title = {{An IoT Endpoint System-on-Chip for Secure and Energy-Efficient Near-Sensor Analytics}},
  journal = {IEEE Trans. Circuits Syst. I Regul. Pap.},
  year = {2017},
  url = {https://openreview.net/forum?id=RGNRZp5CC__},
}
Figure