Profiler shows huge idle time

[alex19999]

I'm new to TF profiler and tensorboard. Could you please tell me how GPU idle time is calculated? I see contradictory data in profiling information on tensorboard:

• I have no blank spaces in events trace (GPU compute stream) presented on trace_viewer page.
• Wherein device (GPU) idle time shown on tensorflow_stats page is about 50%???

It's what I see when I use small batch size. After increasing batch size, same situation is on trace_viewer page, but GPU idle time decreased up to 7%. So, how is idle time calculated?

By the way, another two questions:

1. What is steps on trace_viewer page?
2. Why does trace_viewer page contain only a part of profiling time? I profiled about 100 seconds of execution, but I see only 10 seconds on trace_viewer page.

Thanks a lot in advance!