Hey kano, I'm running 700MHz for around an hour now, without changes to the voltage screw and the NF-A8 PWM running at 2200RPM. Seems stable and good so far. I'm finally getting over 2.5TH/s (that was my goal), so I guess I beat you, with stock voltage! 
Just to make sure: worst that can happen when aiming for too high clock speeds is that it crashes or returns faulty hashes, right? Or is there any 'risk' to pushing the overclock as long as the voltage screw is not touched? And the error rate is displayed through those Java API commands, right? The only way I've checked my error rate so far was the log printed when killing cgminer.
Proof:
2.585Th/s
Well that A says 1.5 hours so definitely higher hash rate than me Just to make sure: worst that can happen when aiming for too high clock speeds is that it crashes or returns faulty hashes, right? Or is there any 'risk' to pushing the overclock as long as the voltage screw is not touched? And the error rate is displayed through those Java API commands, right? The only way I've checked my error rate so far was the log printed when killing cgminer.
Proof:
Code:
(5s):2.647T (1m):2.735T (5m):2.639T (15m):2.602T (avg):2.585Th/s
A:3259855 R:0 HW:1984 WU:36123.2/m
Connected to stratum.kano.is diff 1.62K with stratum as user
Block: 4b258ede... Diff:34.1T Started: [07:05:45.608] Best share: 1.5M
--------------------------------------------------------------------------------
0: GSF 10070017: BM1397:06+ 700.00MHz T:700 P:697 (3:2) | 94.2% WU: 92% | 2.615T / 2.585Th/s WU:36125.3/m
A:3259855 R:0 HW:1984 WU:36123.2/m
Connected to stratum.kano.is diff 1.62K with stratum as user
Block: 4b258ede... Diff:34.1T Started: [07:05:45.608] Best share: 1.5M
--------------------------------------------------------------------------------
0: GSF 10070017: BM1397:06+ 700.00MHz T:700 P:697 (3:2) | 94.2% WU: 92% | 2.615T / 2.585Th/s WU:36125.3/m
2.585Th/s
I wonder if 725MHz is just too high for them (as per my >20% HW) and 700 is the limit?
Also as mentioned in my case, one chips is somewhat of a dud,
so I had to keep pushing the freq higher to get closer to 2.5T,
so for you all 6 chips running OK at 700 will give a higher avg for you also "BM1397:06+ 700.00MHz"
(though my dud does about 25% of 500MHz when you set it to 500MHz)
'Error rate' is HW / (A+R+HW) = 1984 / (3259855 + 0 + 1984)
I changed cgminer to report HW on the screen correctly for the Gekkos - it used to only count HW occurrences, but alas most miners don't return nonces at diff1 - so for the Gekko's it uses the ticket per HW.
In the case of the R909 that's 64 - i.e. 1984 / 64 = 31 bad nonces returned (and each nonce returned is worth 64)
Yes you can kill it - no it's not bullet proof.
It is unlikely to kill it, but as you push it harder that means you are pushing the chips harder also.
If you overload the power it will just switch off, and then the 'fuse' will need a power cycle and letting it cool will fix that.
But the chips themselves, I've no idea under what circumstances they might be at risk.
Maybe sidehack knows?
