GTX 460 superclocked

[Ghost0210]

Just to clarify, EVGA can't report or read a Core clock speed on this card
GPU-z reports the following:
GPU Clock = 608 Mhz
Memory = 802 Mhz
Shader = 1215 Mhz
EVGA reports the same figures as GPU-z omitting the GPU clock figure. nVidia's performance tool reads the same.
But this still leaves the question, where is Boinc reading the 3200Mhz and the 810Mhz from in the stderr text. The effective memory clock of the 465 is rated at 3.2Ghz (4*802mhz), so it still looks to me (as a complete novice in how Boinc gets these figures), that Boinc is reading the wrong clock on this chip as it can't read the correct figure through the API if it is the core clock that is supposed to output to the stderr.
I'm making a lot of assumptions here as I don't have enough knowledge of this to make informed judgements, but it does seem logical to me that Boinc is having issues trying to decipher the API, or the hard coded values

[Jason G]

...for Compute Capability 2.1...

I'd like to know where this figure comes from if you could direct me to that.  I have no Cuda documentation even indicating the existence of this compute capability 2.1, nor how many cores on the die etc.  So someone either has information not available even to registered nVidia early access program developers, NDA or otherwise ... presumably by looking in the drivers/control panel which may not be quite right yet.

I reckon the figure of 48 might be questionable, since that would be 1.5 warps which would require an extra warp scheduler,  or extension of the existing two to support 24 threads each... That isn't a power of two so it wouldn't make sense to make it that way in engineering terms, since hardware engineers only speak binary and hexidecimal ... with the odd exception of some sort of faulty chip harvesting.

I did mix up Ghosts 465, and TouchevGrey's 460 indeed, but the same possibility exists for the libraries reporting, not necessarily hard wired to one value, but broken or generically derived by formula not adapted for 465 or other new stuff.  Also, I think these cards (465 & 460) represent the first the vendors are not forced to use a reference design, but can design their own boards.  That opens the further possibility that the required custom BIOSes aren't all using the same mechanisms as nVidia reference ones, so will probably be the subject of either driver, library or firmware updates as time goes on.

...But this still leaves the question, where is Boinc reading the 3200Mhz and the 810Mhz from in the stderr text....

I still reckon something's broken  [That's the same 810000 shown on my 480... a default of some sort I reckon.  For the 3200 Some addition to the GPU properties retrieved by the library might have shifted the data elements one space in the data structure ... They are kindof nearby, so it'd be a fairly simple task to mess up the properties parsing in the Boinc code ... checking ]

[Ghost0210]

I still reckon something's broken  [That's the same 810000 shown on my 480... a default of some sort I reckon.  For the 3200 Some addition to the GPU properties retrieved by the library might have shifted the data elements one space in the data structure ... They are kindof nearby, so it'd be a fairly simple task to mess up the properties parsing in the Boinc code ... checking ]

Got to agree that something is broken, all the figures indicate (to me atleast) that the 810 and the 3200Mhz figures are either defaults as you said, or Boinc is reading the wrong clock speed (memory instead of core). 
The good news is that this doesn't affect the actual speed of the card, but it would be nice to have the correct figures reported 

[Jason G]

Just looked at (more) Boinc code.  To get the attributes it is using the driver api ( nvcuda.dll ) which is installed along with the driver, instead of the cuda runtime ( cudart...DLL) the app uses.

Probably means the parameters will magically fix themselves with a driver update     I can't see any problems with how Boinc gets the driver API function entry points, nor how it chooses the attributes to get (eliminated the dodgy data structure parsing suspect).

Jason

[Ghost0210]

Good news on that one then. It eliminates any issues with the Boinc source code and hopefully it shouldn't be too much longer till a fresh driver is released onto us

[Jason G]

One step ahead, looking at the schedule for recent 400 series releases:

   GeForce/ION Release 256 WHQL NVIDIA Recommended     258.96     July 19, 2010
   GeForce/ION Release 256 BETA    258.69    June 29, 2010
   GeForce/ION Release 256 WHQL    257.21    June 15, 2010
   GeForce/ION Release 256 BETA    257.15    May 24, 2010
   GeForce Release 197 WHQL    197.75    May 10, 2010
   GeForce Release 197 BETA    197.75    May 3, 2010
   GeForce Release 197 WHQL    197.41    April 9, 2010

*Could be* in a day or so for a new Beta, unless some problem (like this one, lol) causes delays.

[Ghost0210]

Think I cans ee what value Boinc is reading in the stderr now...

I've just been having a play with the shader clocks on the 465 and it would seem that Boinc reads this value and outputs to the stderr for some reason.
I've just changed my shader clock and the corresponding figure is shown in the stderr of a task completed
Here you can see that the shader has been set to 3.2Ghz for some reason (although not showing in any GPU monitoring tools)

setiathome_CUDA: Found 1 CUDA device(s):
  Device 1: GeForce GTX 465, 993 MiB, regsPerBlock 32768
     computeCap 2.0, multiProcs 11
     clockRate = 3200000

After I reset the value back to the correct 1215Mhz

setiathome_CUDA: Found 1 CUDA device(s):
  Device 1: GeForce GTX 465, 993 MiB, regsPerBlock 32768
     computeCap 2.0, multiProcs 11
     clockRate = 1215000

If Boinc only uses the core clock, then why would it output the shader clock to the stderr, unless the bug in the driver is reporting the shader as the core clock?
And with the shader set back to the correct value the Gflops is being reported correctly in Boinc of 855 Gflops:
01/08/2010 16:04:43 |  | NVIDIA GPU 0: GeForce GTX 465 (driver version unknown, CUDA version 3010, compute capability 2.0, 994MB, 855 GFLOPS peak)

[Jason G]

Think I cans ee what value Boinc is reading in the stderr now...
...

that's the app one, [and yes looks wrong].  Stock & my cutdown one use the call from CudaRT.dll  (Cuda Runtime). These device functions in the runtime are just wrappers for the same driver calls, so if one;s broken then they both will be.

[Later:] Moved a bunch of unrelated posts elsewhere.  Nothing to see here, Carry on 

[Ghost0210]

Sorry TouchuvGrey, I kinda hijacked your thread for a while there
Did you manage to find out why your 460 was reporting such a low Gflops rating in Boinc?
From everything I've read about the 460 it should be around the 900 Gflops mark.
With my 465 it seems that the default profile that the nVidia performance tool loaded was corrupt and was shooting my shader clock sky high to give a massively over-r!ted Gflops figure,
Not sure if this would be the same problem (maybe downclocking? as another post said as well)

Ghost

[TouchuvGrey]

Sorry TouchuvGrey, I kinda hijacked your thread for a while there
Did you manage to find out why your 460 was reporting such a low Gflops rating in Boinc?
From everything I've read about the 460 it should be around the 900 Gflops mark.
With my 465 it seems that the default profile that the nVidia performance tool loaded was corrupt and was shooting my shader clock sky high to give a massively over-rated Gflops figure,
Not sure if this would be the same problem (maybe downclocking? as another post said as well)

Ghost

Hello Ghost:

     No Problem at all about "hijacking" my thread. I'm glad that my original post inspired
such an interesting discussion.  i learned some things from the portion of the responses
that i could understand ( i'm not that technical ). i'm still not sure why it is reporting ( still )
only 363 GFLOPS peak. Just a guess on my parts but could the  "compute capability 2.1 "
have something to do with it ? Is BOINC misreading something ? Have i screwed up something
in my computer to the point where the 460 is delivering  far less than it should ?

[Ghost0210]

Hi,

Not sure that you would be able to affect the Gflops that Boinc reports. Richard posted something earlier in the thread about compute 2.1 issues that DA is trying to get an answer from nVidia about. It could be this that is affecting your rating. But Richard/Jason would be the best ones to clarify this issue

Think this is the post from Richard about the compute capability:

And again, this is quite different from the problem with the 460 / GF104 that TouchuvGrey reported at the start of this thread: that one stems from the

'Cuda Cores' @ 48 per Multprocessor ( 7 * 48  = 336 cores


for Compute Capability 2.1

[TouchuvGrey]

Still no answers on what "compute capability 2.1 " indicates.

Another question or so though, SLI enable or disable ?

Given my current setup intel core i7 920 12 gigs of
tri channel DDR3 GTS 250 and GTX 460, what is the best
application for me to run ?

[Ghost0210]

I spent some time a couple of days ago trying to find any information on Cuda's compute capability 2.1 and was unable to find any mention of it, and I haven't seen anything come through on either the Boinc_Alpha list or Boinc_Dev so I would assume that either DA hasn't heard anything back from nVidia about it or he hasn't told published what he was told?

With regards to SLI, common consensus is that it should be disabled for Seti. You should be able to disable it though the Control Centre (I don;'t have SLI enabled cards in at the moment so can't test I'm afraid.

For the application, if you fancy doing some Beta testing for Jason, and you have access to the Beta download area then read this thread:
http://lunatics.kwsn.net/gpu-testing/ongoing-multibeam-cuda-x32f-testing.0.html
The Cuda 3.1 build does have an issue when using mixed cards (on the lesseer card the task will crash), but it would be good if you can confirm on your setup if you experience the same issues
The cuda 3.0 build should work fine on both cards
I have a similar setup to you (GTX465 and either a GT240 or GTX260) so if you could confirm my findings then it would be appreciated.
Please be aware though that these preview applications do have other known issues (no cpu fallback for example) so any issues please report them in the other thread.
If you decide that you want to use these applications and need a hand with the setup, let me know and I'll give you a hand

[MarkJ]

Nvidia seems to use the compute capaility to indicate how many shaders per processor the card has. In the case of the GTX460 (GF104 chip) its 48 per processor.

As for the beta testing, yes I did some on a variety of cards, GTX460 included. I don't mix card types in machines due to driver issues, which have been ongoing with nvidia.

[TouchuvGrey]

i've posted about the GTX 460 over in the nvidia
forum. If i get any answers i'll post them here.