SoyNet, an inference-specialized acceleration solution for AI models, has been developed to replace existing AI frameworks, including TensorFlow, PyTorch, and Caffe, and deliver faster inference services.

Korea Patent Application No. : 10-2018-0136437
PCT International Application No.: PCT/KR2018/013795

AI Inference Accelerator

Intended not for specific AI models but for inference execution, SoyNet can be integrated to run countless AI models.

Faster Execution

According to our tests on the Yolo V3 model, SoyNet’s processing speed was 3 times faster and its memory usage 1/9 less than GPU-based TensorFlow. Performance results may vary depending on specific AI models, but SoyNet, in general, performs 2-5 times faster and uses 1/5-1/9 less memory compared to other accelerators.

Benchmark Environments

Benchmark Environments OS(Windows 10 x64), CPU(i7-8770), Mem(16GB), GPU(GTX1080Ti 11GB)
Model YOLOv3-DarkNet53 (CNN based Object Detection)
Model Size 416x416, Float32, TensorFlow 1.12 (using TensorPT 4.0)

Performance Comparison Video

Play the below video to compare the execution speed of
an AI model on TensorFlow and SoyNet.

Test Environment

GPU : GTX1080Ti, Mddel : Yolov3 416x416 FP32, Accuracy (with the same six significant figures of precision)

Support for Various Models in Combination

The latest deep learning models introduced by recent research articles can only be executed alone and on a high-end GPU like GTX1080ti, SoyNet can run a combination of several models unveiled by the latest studies, ultimately expanding the scope of feasible services.

Lower Cost

SoyNet’s specialized acceleration solutions reduce our clients’ server costs. See below for SoyNet’s cost-cutting benefits in an intelligent CCTV architecture. SoyNet supports over 3 times as many camera channels as before with the same equipment, reducing the cost of installing additional video analysis equipment.

SoyNet's Cost-Cutting Benefits

Intelligent CCTVs are fast emerging as a critical element of a smart city. As the Korean government has successfully completed the Intelligent CCTV Pilot Project, more and more local governments are adopting this innovative security solution. When an AI-based video analysis server is run on SoyNet, it can process more than 3 times more camera channels than before, considerably lowering server costs.

Development Process and Resources Utilization

SoyNet enables you to reduce your reliance on high-cost AI specialists and harness your existing application developers for AI service development. SoyNet lets you solely focus on your domain model in the development phase and run your service right away in its optimal condition once the model is developed.

Development Process

Development Resources Utilization

SoyNet’s Benefits

  • Reduced Costs

    SoyNet supports larger processing tasks using the same equipment, lowering high-end server costs for service execution.
  • Support for High-Speed Processing

    In various environments requiring high-speeds with limited resources, such as autonomous driving, SoyNet offers handy solutions to resolve execution speed issues.
  • Workable with No Internet Connection

    Unlike existing Cloud-based inference processing, SoyNet can be embedded into edge devices, offering inference processing at the edge without Internet connection, thereby supporting businesses to expand their service scope.
  • Greater Value from Existing Developer Resources

    SoyNet APIs enable businesses to develop AI services with their own existing application developer resources instead of highly-trained AI specialists, helping them harness their existing developer pool more efficiently.