NVIDIA vs. Radeon With HITMAN On Linux: CPU Usage, Memory Usage

[theriddick]

Spikes and dips hey? You mean outliner data?

[duby229]

Spikes and dips hey? You mean outliner data?

Those were my words, sorry I don't have the proper vocabulary. The other posts in this thread are fully valid though, I just apparently used incorrect vocabulary to describe what I saw. My bad.

[pal666]

those frame time spikes on fury are bad

[Niarbeht]

One fix could be to check for average value, and remove anything that is above 500% of that (or some other value, just making something up here). Or instead of removing, simply zooming in, and settings the data points to the max value on the y-axis.
Something like this is done in many scientific papers, to avoid focusing on stat fluctuations.

The correct thing to do, likely, would be to set the ymin and ymax on the axes to contain two standard deviations.

[cRaZy-bisCuiT]

Why is the GPU usage graph labeled with "less is better", that doesn't make sense at all.

[Michael]

Why is the GPU usage graph labeled with "less is better", that doesn't make sense at all.

The graphs are all auto-generated and in that case probably need to add a check for GPU usage to display different or probably just drop that string entirely, since in some cases, lower GPU usage could be viewed as better.

[indepe]

How long is the time duration covered by the graphs? And the how long are the intervals between measurements?

It's great that the OpenBenchmarking.org result file contains per-core CPU-usage info. It looks like the port of Hitman does a good job of using all cores. Although in a snapshot-like test, one cannot really tell if these cores are able to do work at the same time.

Unfortunately there are no Windows numbers or much else to compare these results to.

I'm not sure why the OpenBenchmarking.org result file has 5 different GPU-usage graphs, whereas the article has only 2 different ones, plus a redundant copy. I guess in the article, one graph is for the "low" quality setting, and one for the "ultra" setting. The results file has more graphs where even the 1080 GPU maxes out, however so far I couldn't really tell what they belong to.

[franglais125]

The correct thing to do, likely, would be to set the ymin and ymax on the axes to contain two standard deviations.

I thought of something like that too (although I'd personally prefer 3 sigma), but it's easier said than done.

Imagine the following scenario:
- you have essentially three regions, where the data points plateau at 40 FPS, 80 FPS and 200 FPS
- now it's not so clear what "average" and "standard deviation" means, as the proper way to do it would be to analyze each section individually, to identify fluctuations
- following your suggestion might completely remove one of the 3 **legitimate** regions

In the end, as I said, you'd be forced to do a piece-wise statistical analysis, which is already complicating things a tad too much, IMO. This is not a scientific paper, after all.

[franglais125]

Those were my words, sorry I don't have the proper vocabulary. The other posts in this thread are fully valid though, I just apparently used incorrect vocabulary to describe what I saw. My bad.

Your words were fine, nothing to worry about! Scientists use descriptive words also, no need to get academic on simple graphs.

[Ronshere]

I think you would have to incorporate a smoothing function to the graphs however the spikes probably indicate transitions in the game, probably data reading or display processes so I, personally, mentally smooth the charts out and look for the average overall.