x38g reports

[SciManStev]

Here is x39a, x39d, and x39e V6. I have been running 3 wu's at a time with x39a live without issues. I have only had one invalid wu, out of thousands. Driver 275.33

Steve

[Jason G]

Here is x39a, x39d, and x39e V6. I have been running 3 wu's at a time with x39a live without issues. I have only had one invalid wu, out of thousands. Driver 275.33

Steve

Thanks Steve, yeah it's these 460's (and some others) that appear to be sensitive to something I'm abusing.  We seem to be home 'n hosed with the 480s

[SciManStev]

That's what I gathered by reading the threads, but I wanted to throw in a test or two myself. Is there any particular app you would like me to run live, or is there any other comparison you would like me to run?

Steve

PS. I did back down my BCLK one click and eliminated my AP invalids.

[Jason G]

x39e all the way 

[SciManStev]

Done!

Steve

[Jason G]

Manual 27fe11ac.12560.9065.8.10.100  result cross comparison under bench conditions

Claggy's x39e (GTX460) Vs my x39e(GTX480) under bench conditions:  Strongly similar,  Q= 99.95%
My x39e (GTX 480) Vs AKv8bx64SSSE3x:  Result      : Strongly similar,  Q= 99.74%
Claggy's x38g (GTX 460) Vs x39e(My GTX 480): Result      : Strongly similar,  Q= 99.97%
Claggy's x32f (GTX 460) Vs AKv8b( My e8400): Weakly similar. (Bodgy Best Gaussian)
Claggy's x32f (GTX 460) Vs   x39e(My GTX 480): Weakly similar. (Bodgy Best Gaussian)

Tentative analysis:  The known CPU app spikes issues are playing no part here.  The bodgy best Gaussian is in x32f due to innacurate stock nVidia chirp ( ~48 bit precision emulated floating point), I've kept complete documentation on how I fixed that chirp in the alpha ivory beer tower.    x32f & the stock code it came from are crap with highly chirp sensitive signals.

On the live runs, the x38g result likely didn't match the x39d result purely due to the known stability issues we are here to resolve on that class of card.

AKv8b found here:

Spike count:    8
Pulse count:    4
Triplet count:  0
Gaussian count: 2

with the live run CPU 6.03 guy finding:

Spike count:    8
Pulse count:    4
Triplet count:  0
Gaussian count: 3

Now A stock cuda 6.08 wingman has rocked up, STILL INONCLUSIVE....LoL...

Spike count:    8
Pulse count:    4
Triplet count:  0
Gaussian count: 2

IMO, all the results are broken in some way apart from the offline AKv8b ones & the x39d/e results.
Jason

[Claggy]

Here's the results of the 9800GTX+ run, all apps Strongly similar,

Claggy

[Jason G]

Well, unfortunately the 9800 result has the stock-x32f-style bodgy gaussian against AKv8b so there will be something to look at deeper with the pre-Fermis  (That gaussian drift against CPU results has been there for a long time though)

[perryjay]

And here I thought I was just going to show something kinda interesting. I didn't know I was gonna start all this!   

[arkayn]

I thought it was the 560's having problems, I have not seen any problems from my little 460.

[Claggy]

I thought it was the 560's having problems, I have not seen any problems from my little 460.

I think the 560's just have bigger problems,

Claggy

[Jason G]

I thought it was the 560's having problems, I have not seen any problems from my little 460.

I think the 560's just have bigger problems,

and not all 460's & 560's are created equal either, whereas 480's like mine & Steve's are nVidia reference & likely as close to identical as they come, so relatively predictable.  It's something that should not reflect in the results when the code is 100% 'right'. 

Since some of the newer code in play is specifically written to exploit the superscaler instruction level parallelism for maximum bandwidth on compute capabiiity 2.1, i.e 48 instead of 32 Cuda cores per multiprocessor with an extra warp scheduler, if somethings' being pushed to the limits kernel execution configuration wise then It;s going to show on those cards with harder constraints first.

Jason

[Josef W. Segur]

Hmm, I noticed the "Bodgy Best Gaussian" occurred at chirp rate ~63.9958 and the presumably correct one at ~79.3237. So while running the WU with stock 6.95 I periodically checked state.sah to see what would turn up near those rates. Before the first there was a fairly weak "best" captured at chirp -15.88619 with a score of 0.125732, then the bodgy with a score of 1.293129, and the final one had a score of 1.305961. I don't know how to evaluate the 0.98% difference between bodgy and final scores.

AK_v8b_win_SSSE3x.exe has just gotten to bodgy and calculates its score as 1.293091 which is close though I'd rather the first difference were in the 6th significant digit than the 5th.
                                                                  Joe

[Jason G]

AK_v8b_win_SSSE3x.exe has just gotten to bodgy and calculates its score as 1.293091 which is close though I'd rather the first difference were in the 6th significant digit than the 5th.

If you are too, I'm content to call this one "remaining chirp annoyances, with added yet to be divined nVidia Gaussfit implementation vagaries"

[Jason G]

I don't know how to evaluate the 0.98% difference between bodgy and final scores.

Just a naive thought on how that could interact with slight chirp differences:  If you take a fairly 'phat' gaussian (wider bandwidth bin leakage or similar effect) then stride it in slightly different angles, you will get a slightly different fit (shape)  for very similar (but not the same) peak... could there be some substantial aliasing & could [recommending to the project] some windowed transforms improve that SNR (controlling 'lobing' in the frequency domain)?