Performance degradation

[vxf]

I use a GTX550Ti (Fermi) and there has been actual performance improvements with cudaminer more or less until commit 142261bb89144364873ab4c70772a96d16647966 to a nice 82.19 khash/s and just autotuned, unfortunately the latest commits don't do much more than 40-60khash and the autotune makes it even worst.

I kind of isolated the issue to the best commit working here but I don't have more info not even a profiler installed. I may provide more data on request.

[kristianfreeman]

@vxf, just giving my numbers here at the commit you specified compared to HEAD (two Tesla M2090s): I have a funky first GPU which I run a little lower than the second – this brought the first one's hashrate up and the second one's down to average them out around the same speed.

[ouranos]

Same here with a GTX560. Last version I was running was 142261b and running ~126 kh/s. With latest version (827708f) I get ~110 kh/s.

I've reverted to last version and I'm back to ~126 kh/s.

[cbuchner1]

None of you state which kernels you run (F or X). Your cudaminer argument list would be really helpful to understand what commits may have had an impact and why.

2014-01-30 Olivier Brisse notifications@github.com:

Same here with a GTX560 Last version I was running was 142261bhttps://github.com/cbuchner1/CudaMiner/commit/142261band running ~126 kh/s With latest version (827708fhttps://github.com/cbuchner1/CudaMiner/commit/827708f) I get ~110 kh/s

Reverted to last version and back to ~126 kh/s

Reply to this email directly or view it on GitHubhttps://github.com/cbuchner1/CudaMiner/issues/84#issuecomment-33662467 .

[ouranos]

Sorry, my bad: -d 0 -i 0 -q -l F112x2

[kristianfreeman]

-d0 -C2 -lF16x16

[vxf]

@cbuchner1, sorry for the lack of data, I never took the time to actually tune the settings, also I'm not sure if the issue is exactly above 142261bb89144364873ab4c70772a96d16647966 I didn't had the time to test them all, well:

At commit 142261bb89144364873ab4c70772a96d16647966 it autotunes to F32x2 or F16x4 and detects 82.27khash, during the runtime the client says around 78.89 khash/s for both configs and the pools tell 82-85

I now found with -l F32x4 or -l F64x2 it performs the best, the client saying around 79.78 khash/s (yay!!!)

With latest commit and some others in the mean it autotunes to F8x16 detecting 69029.44 hash/s, and during the runtime the client tells 66.79 khash/s

With -l F32x4 at this commit (the best manual tune) the client tells 66.75 khash/s

The X kernel gives much poorer results.

Thank you for your work on this project.

[biozshock]

Have GTX560Ti MSI TwinFrozr. Running with -i 1 -H 1 -l F11x8 -C 1 so i can work at the same time. All were running for 30-40 minutes. One thing with 142261bb89 is that hashrate is more stable, while 02-09 flowing from 100 to 125

	2014-02-04	2014-02-09	142261bb89
speed	~116 khash	~115 khash	~105 khash
temp	93 C	95 C	90 C
(yay!)	20	30	34

And yes, thanks ;)

[vxf]

I consider this now a duplicate of #96 at least for the reasons I started. Continuing now at that issue.