Abstract
Design automation in general, and in particular logic synthesis, can play a key role in enabling the design of application-specific Binarized Neural Networks (BNN). This paper presents the hardware design and synthesis of a purely combinational BNN for ultra-low power near-sensor processing. We leverage the major opportunities raised by BNN models, which consist mostly of logical bit-wise operations and integer counting and comparisons, for pushing ultra-low power deep learning circuits close to the sensor and coupling it with binarized mixed-signal image sensor data. We analyze area, power and energy metrics of BNNs synthesized as combinational networks. Our synthesis results in GlobalFoundries 22nm SOI technology shows a silicon area of 2.61mm2 for implementing a combinational BNN with 32x32 binary input sensor receptive field and weight parameters fixed at design time. This is 2.2x smaller than a synthesized network with re-configurable parameters. With respect to other comparable techniques for deep learning near-sensor processing, our approach features a 10x higher energy efficiency.
Introduction
Deep learning models are notoriously power-hungry, making them difficult to deploy on battery-powered "always-on" devices like smart cameras or IoT sensors. Traditionally, this is addressed by moving computation closer to the sensor (near-sensor processing) to avoid the energy cost of moving raw data. However, even "efficient" models often require significant memory and complex digital controllers. Binarized Neural Networks (BNNs)—which use only 1-bit values for weights and activations—offer a solution by replacing heavy multiplications with simple logical XNOR operations. This research explores the potential of implementing BNNs entirely as raw logic gates, treating the network as a purely combinational hardware circuit rather than software running on a processor.
Methodology
The proposed architecture utilizes a purely combinational hardware design. Unlike traditional accelerators with fetch-decode-execute cycles, the entire network is a single block of logic gates where data flows from input to output in one "ripple."
The core of this architecture is the BinConv Module. This hardware block performs the fundamental BNN operation: XNORing inputs with weights, counting the bits (popcount), and comparing the result to a threshold.
The researchers compared two versions:
- Variable Weights: Allows weights to be changed after manufacturing.
- Fixed Weights: "Hard-wires" weights into the silicon during logic synthesis.
By fixing weights, the synthesis tool can aggressively optimize the circuit. For instance, an XNOR gate with a fixed '1' weight simplifies to a wire, while a '0' becomes an inverter. This enables massive logic sharing and area reduction. For evaluation, models were trained on the MIO-TCD dataset for traffic surveillance. Labeled patches of cars, pedestrians, and cyclists were extracted and resized to 16×16 and 32×32 pixels to fit the small input windows of these ultra-low-power front-end classifiers.
Results and Discussion
The "hard-wiring" approach provides a significant advantage in logic optimization. The synthesis tool identifies redundancies and reuses partial results across the network, a feat difficult to achieve in software.
Key performance metrics include:
- Area Efficiency: A 32×32 input BNN with fixed weights occupies only 2.61 mm² in 22nm SOI technology, which is 2.2× smaller than the reconfigurable version.
- Energy Efficiency: The approach achieves approximately 450-670 TOp/J, roughly 10× higher than YodaNN, a leading binary weight accelerator.
- Performance: Theoretical throughput reaches 50 million frames per second, though practical limits are set by the sensor frame rate.
- Accuracy: The 32×32 model achieved 64.7% accuracy on the MIO-TCD task. While lower than cloud-based models, it serves as an effective "early trigger" to wake up more powerful systems upon detecting relevant events.
A notable bottleneck identified is the Electronic Design Automation (EDA) tools. Current tools struggle to synthesize networks larger than 64×64 due to the exponential logic complexity, which causes software memory exhaustion.
Conclusion
Implementing BNNs as purely combinational circuits offers a "quantum leap" in energy efficiency for near-sensor processing. While the fixed-weight approach sacrifices flexibility, it is ideal for stable, specific tasks like human detection in "always-on" sensors. Future work must address scalability by developing "BNN-aware" synthesis tools and expanding the design to handle multi-channel or higher bit-depth inputs.
Summary updated