<aside> ⚙

Plan Structure

Context → Aim → Objectives → Principles → Documentation → Tests

</aside>

TL:DR - We plan to run real-world tests examining how model quantisation affects energy consumption and performance on the Nvidia Jetson Orin. We will verify and validate the Jetson’s onboard ****power/energy readings by comparing them to measurements from a third-party external device and an external battery. This ensures our subsequent quantisation impact tests have reliable power data.

1. Context

We want to understand the comparison between real-world performance and energy usage when running quantised and non quantised models on the Jetson Orin Nano. However, before drawing any conclusions about quantisation’s energy gains or trade-offs, we must confirm the accuracy of the Jetson’s own power/energy readings. For this, we will compare onboard measurements to readings from a third-party measurement device.

2. Aim

Test how a quantisation model affects performance and energy consumption in real-world examples (object recognition) compared to a non-quantised model

3. Objectives

• Set up a test environment using real-world images where the Jetson Orin Nano can run both quantised and non-quantised models.
• Verify the Jetsons onboard metrics by checking power/energy readings with a third-party external device.
• Use an object recognition model such as YOLO as the baseline values
• Compare energy consumption with performance metrics ****(throughput, latency, etc) under different quantisation levels.
• Identify ****anomalies or measurement issues between onboard and external readings.
• Document all findings