Here's where that came from, ckolivas's recommendation when he released the S3 binary I am currently using on the OC unit:
Here's an updated S3 binary.
http://ck.kolivas.org/apps/cgminer/antminer/s3/4.6.1-141020/cgminerRecommended if you're mining on p2pool for the default binary actually discards stale shares which you should never do, especially on p2pool. Also includes changes to queuing and memory usage that were necessary on S4 but probably only of minor benefit here. Recommend you edit the cgminer startup script to remove the --queue value entirely, and add --lowmem. Performance should be pretty much unchanged.
Death spiral on the OC unit definitive now, average down to 523 GH/s. All chips good with "o", still just the 3 HW errors, temps 42/38. This would bother me less if I had some explanation for what is happening.
On the queue, I'll take ckolivas' word over what I may recall ...
And yes, that is more like a death spiral, too damn right! I'd have expected it to be flashing x's all over + an increase in HW errors at this point, but then again the new binaries support that --bitmain-hwerror option (or such like) that I have never gotten my head around! I think its that time to try the 0800 setting .... I am convinced it is a heat problem your rigs are encountering due to its consistency in drop-off, so reducing the voltage may help (but may need a power cycle, infact I'd say do one even though I do all my tests initially without one).