Sorry, there's a flaw in the 1.18 OpenCL code that let (only) the OSX ATI OpenCL driver error out. Usually we find such errors ourselves, but this in a part of the code that requires double precision support, which our card doesn't have. I'll try to get this fixed and publish a new version later today. For now the 1.18 for OSX has been deprecated.
FWIW I just shuffled our graphics cards such that we now have a double precision ATI card in our 10.9 Mac Pro, so we should be able to find such problems next time and fix these before release.
Wow, processing time on a 7870 on Windows is cut by over 30%, even though a 7870 doesn't have very good double precision performance as far as I know. Great work!
In the beginning of 1.18 i had also a lot of failures;
Increased cooling & slow down the core clock from 100% to about 70% (R9 290).
At this Moment, only 1 failed WU since 14 hours. (before i speed down, 18 WUs failed of 1.18 beta)
GPU-Z increased GPU load from about 30% to 55%.
have the logs if needed.
System run 24h.
Have also the feeling, that clock decreasing & better cooling is the best solution. Throttling is always on 65°C
(1.18) 1250 sec completing a WU -> 100% clock speed
(1.18) 1310 sec completing a WU -> 70% clock speed
decreasing the clock Speed slow down the completion in this case 5%, but i have valid results :-)
comparison to
(1.17) 2800 sec completing a WU -> 100% clock speed
everytime running 2 WU's at once.
Great improvement the 1.18, thanks to Christophe for using his holidays for this; let's fix the failures.
My Xeon-based crunching-only machine has managed to get through 60 of the 1.18's, so I thought I'd examine performance so far. I'm also including figures for my GTX 650. The GTX 650 machine disables GPU computing while in use, and the machine got used a lot this past holiday weekend, so it hasn't had a chance to work through many 1.18's. Very similar performance differences.
Bernd Machenschalk wrote:The
)
Thanks Bernd for the precision.
Sorry, there's a flaw in the
)
Sorry, there's a flaw in the 1.18 OpenCL code that let (only) the OSX ATI OpenCL driver error out. Usually we find such errors ourselves, but this in a part of the code that requires double precision support, which our card doesn't have. I'll try to get this fixed and publish a new version later today. For now the 1.18 for OSX has been deprecated.
BM
1.19 Beta released (OSX only)
)
1.19 Beta released (OSX only)
BM
Bernd, great news -
)
Bernd,
great news - vielen Dank!
FWIW I just shuffled our
)
FWIW I just shuffled our graphics cards such that we now have a double precision ATI card in our 10.9 Mac Pro, so we should be able to find such problems next time and fix these before release.
BM
Wow, processing time on a
)
Wow, processing time on a 7870 on Windows is cut by over 30%, even though a 7870 doesn't have very good double precision performance as far as I know. Great work!
Matt_145 wrote:1.18 is a
)
I can confirm this for my GTX 1070. Running 3 tasks at a time, completion times went from around 42 minutes to around 28 minutes.
Proud member of SETI.Germany
Bernd Machenschalk wrote:1.19
)
Hi Bernd,
sorry - bad news - these are bomming out on my machine also. Please see here; https://einsteinathome.org/host/12464084/tasks/error
Let us know when there's another new beta to try, I'll run the old 1.17s for now.
Cheerio,
Kailee.
In the beginning of 1.18 i
)
In the beginning of 1.18 i had also a lot of failures;
Increased cooling & slow down the core clock from 100% to about 70% (R9 290).
At this Moment, only 1 failed WU since 14 hours. (before i speed down, 18 WUs failed of 1.18 beta)
GPU-Z increased GPU load from about 30% to 55%.
have the logs if needed.
System run 24h.
Have also the feeling, that clock decreasing & better cooling is the best solution. Throttling is always on 65°C
(1.18) 1250 sec completing a WU -> 100% clock speed
(1.18) 1310 sec completing a WU -> 70% clock speed
decreasing the clock Speed slow down the completion in this case 5%, but i have valid results :-)
comparison to
(1.17) 2800 sec completing a WU -> 100% clock speed
everytime running 2 WU's at once.
Great improvement the 1.18, thanks to Christophe for using his holidays for this; let's fix the failures.
My Xeon-based crunching-only
)
My Xeon-based crunching-only machine has managed to get through 60 of the 1.18's, so I thought I'd examine performance so far. I'm also including figures for my GTX 650. The GTX 650 machine disables GPU computing while in use, and the machine got used a lot this past holiday weekend, so it hasn't had a chance to work through many 1.18's. Very similar performance differences.