Interesting, thanks for the info.
I don’t have a 65F ambient so I have to work with 70F+… except in winter.
Each of my rigs has a secondary box fan for supplemental air flow. This is more to direct the heat toward the cooling system intakes but also cooled the GPU’s a few C.
For me: I found that you must maintain a ratio between the GPU OC and Memory OC. There is a pretty significant difference when you have the right ratio (just like a CPU). I have code that sweeps the Mem OC for each GPU OC on each card. Once I have identified the Max GPU OC then the Mem OC can be set to the optimum. Once I am at this point the temp is the rate limiting factor for my H/s.
I set my GPU’s at 62-63C because I discovered that once the OC is optimized the GPU’s will start to become unstable with power and temp. For example: when my GPU’s were set to 62C and the winter temps plummeted the software tried to increase power to maintain a 62C temp. However, several GPU’s on different rigs started to become unstable. I didn’t expect this and my software did not have a power limit set for reboots (they do now). Then (again cold) I had GPU’s start to drop hash rate (not hang just drop hash). The software compensated with increased power. At some point the power could not compensate for the lower temps and the rig rebooted due to lower GPU performance (I allow 10% drop from the GPU moving average). However, the power was now set rather high, and once rebooting the temps hit my limit of 70C and caused a reboot before the power level could be dropped (reboot loop). That was harder to fix but my rigs text me when they reboot more than 3 times in 15 min so I caught it pretty quickly. Now I sweep power setting as well after GPU OC and Mem OC have been set. This gives me a power limit for each GPU as well.
So my point being that I have run into more issues with cold rather than heat so water cooling would not increase my earnings since I am running cold now. Running the fan at 100% on my rigs ensures I get the max hash rate I can with the absent temp and a power limits set for each GPU. However, in all honestly I have not tried a fan sweep… but that would be a very difficult matrix to interpret.
There is also a harmonic present on Nvidia GPU’s between the GPU OC and the Mem OC. Its rather hard to find but its ~20H/s, so worth it. I have just started writing beta code to find the oscillation and compensate for it.
Would be interesting to compare notes sometime.