980 Ti SLI Overclocking question

[titan_rw]

Hi all.

I have two 980 Ti's.  They're EVGA FTW ACX cards (non reference).  I have Kraken G10's on both, with a Kraken G61 on one (gpu1), and a G41 on the other (gpu2).  I left the backplate / midplate on, and am using a copper shim to make contact between GPU core and heatsink / block.  I also put some additional heatsinks on the midplate directly above the VRM's more good measure.  VRM temperature (measured by touching the midplate) is good. (much better than under stock air cooling, even with the fans forced to 100%).  Water temps are as displayed by CAM.

 

I figured I'd try overclocking them now, as temperatures are extremely good.

 

Here's my stock temps / clocks / voltages under furmark, and cudahashcat:

Ambient temperature 23C.

 

Furmark:
1430 / 1392 core
1.199 / 1.187 V
3506 3506 mem
39 / 47 C gpu
40 / 45 C water

 

CudaHashCat:
1418 / 1380 core
1.180 / 1.187 V
3305 / 3305 mem
42 / 49 C gpu
42 / 45 C water

Speed.GPU.#1...: 15625.7 MH/s
Speed.GPU.#2...: 15248.3 MH/s

I played around with OC settings, and found that +580 memory clock was the highest I could get.  I backed it off to +400 as the core clock was giving me trouble, and I understand that the memory clock doesn't result in that much gain.

 

Here's what I tried: +87 mV, 110% power limit, +50 core clock, +400 memory clock

 

CudaHashCat: (rock solid, showing 105% to 107% power usage).
1480 / 1455 core
1.218 / 1.224 V
3305 / 3305 mem
44 / 53 C gpu
43 / 48 C water

Speed.GPU.#2...: 16180.2 MH/s
Speed.GPU.#1...: 16148.9 MH/s

 

Furmark: (showing 97% to 100% power usage, but fails about 30 secs in, GPU 2 core clock drops to 595 Mhz.
frame rate drops.  Temperatures were OK before the cut in core speed.)

 

Verified reduced performance on GPU2 in hashcat:
Speed.GPU.#1...: 16628.4 MH/s
Speed.GPU.#2...:  7078.4 MH/s

 

Only a reboot will restore the gpu core clock performance.

 

Everything was stable under CudaHashCat, which uses more power, and produces more heat than furmark.

Is furmark over sensitive about locking down core clocks?

 

I was thinking I'd be able to get more than +50 mhz core OC with the cards water cooled, even on AIO's.

Should I ignore the furmark failure, and rely on temperatures in cudahashcat?  It runs hotter than furmark anyway.

Thanks in advance.