Well, not exactly. Our tests didn't exhibit these accuracy issues and they do occur only on a certain fraction of systems.
Well, that is, what I can confirm. I have 4 computers running
3 with 64 bit WIN7 (2 with 64 bit boinc 6.10.58, 1 with 32 bit boinc 6.10.58)
1 with 32 bit Vista
The three win7 machines do have the problem
The 32 bit Vista machine always has valid results (except one with validation error), when the wingman to CUDA is SSE
We know which things have an impact on this and we'll be working on a fix very soon. However, we might depend on your feedback in order to verify that those fixes indeed improved the situation since we probably can't test/reproduce the observed behavior using our own test systems.
Thanks,
Oliver
In that case I will keep an eye on my results and will post any more "Inconclusive results" I come across.
I find on my Win7 machines that trying to pump a vid card a bit much will cause this too. Instead of just crash and burn, win 7 will momentarily 'lock up' reset the video card driver, and adjust the clock speeds WAY down on the card, and pick up where they left off at (almost). When this happened, almost without fail, any GPU tasks I was doing would end up with a computational error or invalid in the validation. While the settings may be well within what your card can handle on other apps, some of them do not like speeds set high. I have also seen where even at a certain setting, some runs will work, others will fail.. like it's on the threshold there.
As an example milkyway I can fire the card up a lot,
einstein I better be moderate
gpugrid I better be very gentle or it will die off quickly.
Try adjusting your vid and vid memory speeds down a tad and see if this helps.
Aaron
If god meant for us not to BOINC he'd have made our #$%^%^ shorter!!
This is a known problem that they have not fixed yet, sometimes you get 2 cpu's that validate each other, other times you get gpu's that validate each other. When yours is the odd man out, you get NO credits for that unit!
This is a known problem that they have not fixed yet, sometimes you get 2 cpu's that validate each other, other times you get gpu's that validate each other. When yours is the odd man out, you get NO credits for that unit!
And sometimes Cuda and SSE simply validate each other with no problem whatsoever.
Homogenous redundancy would probably help. Rounding problems are really mean, even worse when MSVC libraries are involved - different versions have different rounding policies (bad experiences from the RNAWorld development phase) :-/
A huge project like this one could easily affort enabling HR.
RE: Well, not exactly. Our
)
Well, that is, what I can confirm. I have 4 computers running
3 with 64 bit WIN7 (2 with 64 bit boinc 6.10.58, 1 with 32 bit boinc 6.10.58)
1 with 32 bit Vista
The three win7 machines do have the problem
The 32 bit Vista machine always has valid results (except one with validation error), when the wingman to CUDA is SSE
RE: We know which things
)
In that case I will keep an eye on my results and will post any more "Inconclusive results" I come across.
We have enough examples for
)
We have enough examples for the time being. You may wait until an updated version is released.
Thanks,
Oliver
Einstein@Home Project
I find on my Win7 machines
)
I find on my Win7 machines that trying to pump a vid card a bit much will cause this too. Instead of just crash and burn, win 7 will momentarily 'lock up' reset the video card driver, and adjust the clock speeds WAY down on the card, and pick up where they left off at (almost). When this happened, almost without fail, any GPU tasks I was doing would end up with a computational error or invalid in the validation. While the settings may be well within what your card can handle on other apps, some of them do not like speeds set high. I have also seen where even at a certain setting, some runs will work, others will fail.. like it's on the threshold there.
As an example milkyway I can fire the card up a lot,
einstein I better be moderate
gpugrid I better be very gentle or it will die off quickly.
Try adjusting your vid and vid memory speeds down a tad and see if this helps.
Aaron
If god meant for us not to BOINC he'd have made our #$%^%^ shorter!!
When my W7 machine locked,
)
When my W7 machine locked, and down clocked the card, it was the Nvidia drivers crashing, all seems O.K. now using V266.35.
My GTX 460 is a STD. OC model, so should be fine at it's rated speed.
More investigation needed IMHO.
dunx
I posted it in the wrong
)
I posted it in the wrong thread, so here's my post in the right one:
I fail to see a pattern with validation/inconclusiveness.
Edith says:
Most actual inconclusives are against SSE, but one against an ordinary 1.04.
I have some valid agains SSE.
I had valid 1.04 against Cuda.
Grüße vom Sänger
same here ... repost from the
)
same here ... repost from the "validate errors" thread :
BRP3SSE and CPU vs. BRP3cuda32
wuid=91736864 (cuda won)
wuid=91723581 (CPU won)
RE: same here ... repost
)
This is a known problem that they have not fixed yet, sometimes you get 2 cpu's that validate each other, other times you get gpu's that validate each other. When yours is the odd man out, you get NO credits for that unit!
RE: This is a known problem
)
And sometimes Cuda and SSE simply validate each other with no problem whatsoever.
Grüße vom Sänger
Homogenous redundancy would
)
Homogenous redundancy would probably help. Rounding problems are really mean, even worse when MSVC libraries are involved - different versions have different rounding policies (bad experiences from the RNAWorld development phase) :-/
A huge project like this one could easily affort enabling HR.