Keeping people informed, here is the last response I got from the Sapphire tech I'm corresponding with and my reply. Hopefully we will all start to get these issues resolved now.

Hi Dave :

Thanks a lot for your update.

I will need your help about some information that we will like to duplicate this issue at our side.

Actually , we still not got related information about this kind of testing would crash the board.



1. What applications or 3D games that you used to test the card ?

2. How to check the VRM usage during 1200~1300MB , use GPU-Z ? (We only saw memory usage , is this what you meant ?) any photos can provide to us ?

3. Please show related photos about how to adjusting the memory voltage with TriXX V4.4.0b

4. Please provide us your detail platform information included OS and PSU.

5. Can you try to set to V1.55V at your side and try again if will crash or not ?



Thanks a lot for your help.
My reply:

1. To maximise vRAM usage I have tested with Skyrim using the HD2K texture mod (http://skyrim.nexusmods.com/mods/607) as well as my own HD textures for Stalker: Shadow of Chernobyl (http://www.moddb.com/mods/stalker-twilight-ket-2zero) both games crash when vRAM usage goes above 1200MB - 1300MB. Sometimes it varies a little and won't immediately crash, but it never takes long once vRAM usage gets to the 1200-1300MB region. For clarification, I have tried using Catalyst 12.10, 12.11 Beta11, 13.1 and 13.2 beta drivers, each time doing a thorough cleanup after the normal Catalyst uninstall process with atiman uninstaller 7.0.2 and Driver Fusion 1.1. This problem is most definitely because memory voltage is too low as like I have said, setting the memory voltage to 1.6v completely fixes this problem and I have also used a MSI HD7950 TwinFrozr and do not experience this problem, its memory voltage by default is 1.6v and also uses Hynix memory.

2. Yes I monitor vRAM usage with GPU-Z 0.6.7 - Memory Usage (Deadicated) sensor reading.

3. Attached is a copy of the modified version of Trixx 4.4.0b. This version allows for adjustment of memory voltage, the standard version of Trixx 4.4.0b does not. Also included is a picture showing the modded version of Trixx with memory voltage set to 1.6v, also with GPU-Z verifying the voltage is actually set and also showing the card with a vRAM load over 1600MB running fine (GPU load 99%).

4. My full system spec:

intel Core i5 2500k
Asrock Z77 Extreme6 w\ UEFI 2.6
2x4GB G.Skill Ripjaws PC12800 7-8-7-24
640GB Western Digital Caviar Black, SATA2 32MB Cache
2x SATA Sony Optiarc DVD-RW
Asus Xonar DX 7.1 (unixonar drivers)
Corsair HX 850w
Windows 7 Ultimate SP1
All latest drivers for chipset, graphics, sound, etc

5. I have tried setting 1.55v for the memory but stability is patchy, not as bad, but still not 100% stable. The absolute minimum memory voltage I can set that is seemingly stable is 1.58v, but I set memory voltage to 1.6v to factor in such things as the possibility of memory vdroop. Both Elpida and Hynix (Vapor-X I have uses Hynix memory) verify that at higher memory frequencies such as what the Vapor-X runs at memory voltage needs to be in the region of 1.55v - 1.6v, depending on the application of the memory and choice of components for the design. You can see this for yourself here;

Elpida; http://www.elpida.com/en/products/gddr5.html
Hynix; https://www.skhynix.com/products/gra...No=H5GQ2H24MFR

I have also tested GPU vdroop levels, when GPU vdroop levels were tested Heaven 3.0 was used with the following settings (Heaven doesn't use as much memory as the other tests so the card doesn't crash):

Tessellation: Extreme
Shaders: High
AA: 8x
AF: 16x
Resolution: 1680*1050

Board Li power was set to +20 then measurements of GPU vdroop were taken with GPU-Z 0.6.7 while Heaven 3.0 was left to loop. I also let a colleague at work measure GPU vdroop directly via hardware as he works more hands-on with the actual hardware side of things than I do. His findings were similar to what GPU-Z reported, vdroop is anywhere in the region of 0.025v to 0.078v depending on the level of GPU usage and the level of AA / AF. With a 100% GPU utilisation, Extreme Tessellation and 8x AA, 16x AF applied GPU vdroop was as bad as 0.078v and never better than 0.050v.

I've also had to completely redesign the VRM cooling on the Vapor-X, at stock frequencies with Heaven 3.0 using the above settings VRM temps hit in excess of 126c very rapidly, suffice to say and unsurprisingly the card first started having wild GPU frequency fluctuations (constantly bouncing between 501MHz and 950MHz) then soon after the card crashed with VRM temps that high. During this period GPU temps were never hotter than 65c (measured with a infrared temp gun thats accurate to +/- 2%)

Regards, Dave