Accelerating Performance, Power and Functional validation of Computer Vision Use cases on next generation Edge Inferencing Products

In this paper we present the methodology to accelerate three validation vectors targeted for Computer Vision use cases on AI edge inferencing products using Emulation platform. These are Functional, Performance & Power validation. Vision is one of the key subsystems for Edge Inferencing AI products to run complex use cases for Neural Networks and Compute Algorithms. Along with Multiple camera inputs, multi-stream Media encode/decode capability, Graphics processing, dual display support, the bring up of end to end functional use cases in a pre-silicon environment poses several challenges. The next step, Performance & Power analysis for these use cases to ensure right architectural and design tweaks is of paramount importance. We chose Emulation as the pre-silicon platform to validate end to end Neural Network use cases such as RESNET50, RESNET101, Mobile-NET, Tiny Yolo for object detection, tracking and image classification. The proposed methodology leverages IP/Subsystem environment at SOC level to generate not only the testbench but test vectors as well. Building on top of it, Performance validation framework utilizes these High resolution, Multi-frame Firmware based scenarios for assessing the system level performance metrics with close to silicon accuracy. The environment thus put up is leveraged for the 3rd validation vector – Estimation of average and peak power consumption as well.

Yoga Priya Vadivelu, Intel Technology India Pvt Ltd
Arpan Shah, Intel Technology India Pvt Ltd
Deepinder Singh Mohoora, Intel Technology India Pvt Ltd
Ullas Piyush kanti karmaka, Intel Technology India Pvt Ltd
Praveen Buddireddy, Intel Technology India Pvt Ltd