Dudaji, a Korean company that builds custom computer vision and natural language processing solutions for businesses, is building new services that use Furiosa's WARBOY accelerator for highly efficient inference at scale. As part of that work, they evaluated how well WARBOY will work for demanding applications such as monitoring workplace safety, where the AI system must quickly detect complex issues such as an employee not using protective equipment or driving dangerously on a worksite.

Dudaji successfully ported a range of custom YOLO object detection models to WARBOY and then optimized them through calibration and quantization. The end result achieved 40 FPS on eight concurrent streams using a YOLOv5s model for tasks including fire detection, worker detection, safety gear detection and construction site monitoring. 

The models were tested for performance and also for efficiency, which is critical in real-time monitoring situations where rapid and reliable inference is necessary. Peak performance was 330FPS in total on a single WARBOY card.

“The model performs excellently in actual environments, detecting fires accurately and contributing to the prevention of fire damage,” said Dudaji’s Joonwon Lee. z

End to end deployment

Dudaji began by assessing WARBOY’s model zoo. In evaluating a potential AI hardware solution, it’s crucial that the device supports the algorithms that will work best for a variety of applications. The YOLO family of models has become an industry standard because of its ability to perform real-time object detection in a single pass, making it much faster than traditional CNNs. WARBOY supports YOLOv5m, YOLOv5L, and YOLOv7w6Pose (for pose estimation).

WARBOY is available on Kakao Cloud, so Dudaji was able to easily access the WARBOY instance remotely and test the Furiosa SDK's compatibility with public models and custom algorithms. They confirmed WARBOY’s SDK was compatible with major AI frameworks like PyTorch and TensorFlow Lite.

The WARBOY SDK allows engineers to easily convert Torch models to the ONNX format, so they can then calibrate and quantize using ImageNet samples.

Optimization

Dudaji performed distributed parallel execution of experiments to select models based on quantitative metrics (time accuracy) and various quantization algorithms (e.g., MIN_MAX_ASYM, ENTROPY_ASYM) suitable for different data sets. They found MinMax calibration was most effective.

In testing the YOLOv5s model, Dudaji found preprocessing took between 4 and 10 minutes (including model optimization, calibration, quantization), and inference operated at approximately 40 frames per second (FPS).

Dudaji reported high levels of operational stability with WARBOY, even under continuous operation and in scenarios simulating overload conditions. The company’s engineers assessed WARBOY’s stability during prolonged continuous operation and in overload situations and found there were no hardware faults or other errors over the project’s six-month test period. This reliability is crucial for ensuring uninterrupted safety monitoring in critical work environments.

Dudaji is now preparing to work with FuriosaAI to port additional applications (e.g., detecting copyright infringement and language translation) to Furiosa’s second-gen chip, RNGD, which will launch later in 2024.