Forum > GPU crunching

GTX 460 superclocked

<< < (3/23) > >>

Ghost0210:
Thanks all
From what I've got above it does seem to make sense.
This fits in niceley with what I'm seeing, when I first boot the box, Boinc gives a Gflops of 570, add 50% on to this estimate from Richards post and this gives the correct Gflop rating of 855 but this is a GF100 chip so not sure why you would have to do this? Then, using nVidia's control panel - change the fan speed to manual and up it to 65% and restart Boinc it now gives the 2.5 Tflop rating.
I have a feeling that this may have to do with the way that Boinc reports the clock rate, EVGA Precsion and GPU-Z all rate the clock at 810 Mhz and if I don't use nVidia's toll to increase the fan speed, Boinc will report that 810Mhz clock rate in the stderr of any tasks.
But, when the Gflops is being reported as 2.5Tflops the clock rate in the stderr is being reported as 3200Mhz.
Since this figure has really no effect on the speed of the tasks or the performance of the card, like I said I'm not really bothered what Boinc reports it's more out of personel interest than anything

Richard Haselgrove:
If we can track down what's happening, I can add it to my _alpha post about the shader count - that needs a bump, in any event.

Remember that CUDA cards have three different clocks. The only one that matters for BOINC is the shader clock, which would be something like 1.215 GHz for the 465. If you're seeing 3.2 GHz, then it looks as if both BOINC and the app are picking up the memory clock instead - that would be a bug. But whether it's a bug in the app(s) coding, the NVidia API, or EVGA Precsion, I couldn't begin to tell. Jason?

Ghost0210:
EVGA Precision and GPU-Z are both picking up 1215 Mhz for the shader clock which is correct, but if you look at the stderr for a task completed on the 465 when it is reporting a Gflop of 2253 then the clock rate is being picked up as 3.2Ghz


--- Quote ---<core_client_version>6.11.4</core_client_version>
<![CDATA[
<stderr_txt>
Before cudaAcc_initializeDevice(): Boinc passed gCUDADevPref 1
setiathome_CUDA: Found 1 CUDA device(s):
  Device 1: GeForce GTX 465, 993 MiB, regsPerBlock 32768
     computeCap 2.0, multiProcs 11
     clockRate = 3200000

--- End quote ---

I've attached the full stderr for you, I'm assuming that the clock rate in the stderr is the shader clock?
If you look at a task that has been completed with a gflop of 570, then it picks up the 810Mhz clock rate, which according to GPU-Z/EVGA is the Memory clock


--- Quote ---But whether it's a bug in the app(s) coding, the NVidia API, or EVGA Precsion, I couldn't begin to tell.
--- End quote ---
Don't think its a bug in EVGA as GPU-z is reporting this as well as nVidia's Performance tool

Jason G:
Most likely a driver API or Cuda Runtime library bug/omission.  Not surprising since it's a new card & there are almost always some teething issues with drivers & new card models. By my understanding GTX 460 is a new die (GF104), rather than a cut down GF100, correct me if I am wrong.  That could mean the clock circuitry &/or strategies are entirely different.  If that's the case, it's probably due to refinement of the memory controller, which was proving to be the limit for GF100.  Reports I've had from a Gamer/Benchmarker  friend seems to indicate that something like that has indeed been 'fixed' and achieves 'Unexpectedly high performance' for the money.

To clarify in case of any confusion: Boinc, for its peak flops estimates should be using the 'Core Clock' specification, which on my 480 is currently at 782MHz, but shows in my task stderr output as '810000', so I feel it's a generic driver or Cuda library encoded figure rather than necessarily read from the current rate.

Fermi's have the shaders locked at 2x the core clock (i.e. 1564MHz on my 480) ... but we don't use shaders, we use 'Cuda Cores' @ 32 per Multprocessor ( 15 * 32 = 480 cores )

- Boinc shows mine as 778GFlops Peak
- Using actual figures: (1000.*782000) * 15 * 32 * 2 = 750.72 GFlops peak
- Using figures from stderr: (1000.*810000) * 15 * 32 * 2 = 777.6 GFlops peak

So Boinc is using the 'slightly dodgy' figure reported by the Cuda Runtime, via the cudaGetDeviceProperties() call as the app does. Since mine is reporting a generic figure, slightly overstated even though it's a reference card, we may not see 'real; figures on newer cards until a new Cuda SDK & library version comes out.

Richard Haselgrove:
But Ghost's 465, although a new-ish card (released 31 May), has the older GF100 chip. Should have got the driver bugs out of that by now.

And he reported earlier in this thread that "[BOINC GFlops] can go from 570->855->2553 Gflops in a day". That doesn't sound like a static library value: more like idle/downclocked: then running normally: then a borked figure from EVGA Precision. In general, I don't buy the library idea (although Jord seems to like it too): people who over/underclock these cards report BOINC reported speeds which vary in proportion, so I think the runtime is genuinely getting some representation of a clock speed, even if not always the right one. :'(

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

Navigation

[0] Message Index

[#] Next page

[*] Previous page