Efficient Frontier Methodology

To build the Efficient Frontier, we uploaded SOTA computer vision models, compiled them for runtime, using deci runtime optimization, and benchmarked on different hardwares. To perform the benchmark we streamed random noise input data for 60 seconds, and checked the total latency including: copying the data in both directions (host to GPU and GPU to host) and running a forward pass through the model. Please see more technical details below.

Classification Frontier:

• image resolution: (3, 224, 224)

• dataset: Imagenet

• accuracy metric: Top-1

Object Detection Frontier:

• image resolution:

For YOLOX tiny, YOLOX nano, DeciDet2 - (3, 416, 416)
For YOLOX small, YOLOX medium; YOLOv5, DeciDet5, DeciDet4, Decidet3 - (3, 640, 640)
For DeciDet1 - (3, 320, 320)

• dataset: COCO
• accuracy metric: [email protected]:0.95