Timing a perception pipeline: where do you start the clock?

OscarAbbott0010 · 2026年4月7日 10:53 UTC

已关闭

主帖中文翻译暂不可用,以下显示英文原文;回复保留原文。 查看英文原文

OS
OscarAbbott0010
Jetson AGX Thor developer kit pictured in NVIDIA's availability announcement
August 2025 developer-kit starting price: US$3,499; the volume module price is a different offer. Product image: NVIDIA. Published 2025-08-25. Source

Back to my Thor question from August. I'm defining the benchmark before deciding whether to buy anything. If I time from when software receives a frame, I might miss time spent inside the camera. What timestamps would you want in a robot-perception report? The old launch link is context for the hardware choice, not a benchmark result.

https://blogs.nvidia.com/blog/jetson-thor-physical-ai-edge/

3 条回复

PR
PriyaIO

Write down what each timestamp actually represents: exposure if available, receive, preprocessing complete, inference complete, command issued. Don't pretend clocks on separate devices are aligned unless you've checked that.

0分
RI
RickShift

Keep the slow cases. Average timing is lovely right up to the one pause that matters.

0分
OS
OscarAbbott0010

Thanks both. I'll show the distribution and describe the clocks, not just one average. Much clearer question now.

0分

讨论已关闭

此讨论已连续六个月没有活动,现已关闭新回复。 最后活动: .