A reminder for Curve Optimizer testing: Corecycler comes with a config.ini for a reason. Probably this was already mentioned in the CO thread of biblical proportions, but I'm a smooth brain that only found out last week.
Obviously, any all-core benchmark is a terrible test (by itself, at least) for Curve Optimizer, PB clocks drop like a stone on all-core. But default settings Corecycler doesn't seem to be too great either.
By default, the script only runs 6 minutes per core, at "Huge" FFT length (basically just the region beyond the Large FFT sizes). 6 minutes is not nearly enough to test all members of even the Huge FFT region, it's only enough roughly for a full run of Smallest or Small FFTs (which the script does not test by default).
I had long been running a -10/10/12/12/12/12/15/15/15/15/15/15 profile, which I assumed to be stable due to always passing default 6 min Corecycler and generally no symptoms of instability.
- I started by changing to 6 min Small FFT (~1 full cycle per iteration), then 13 min Small FFT (~2 cycles). The preferred cores started dropping like flies, rounding errors for days.
- I tested 13 min Huge FFT (~1 cycle) next. Having been adjusted, the cores were much closer to stable, but some threw sporadic errors that didn't necessarily manifest in every test iteration.
- Lastly, 68 min All FFT (1 cycle). Obviously *very* time consuming, but it continued to discover cores that needed further tweaking.
If I need to start a specific core from scratch, I start with 13 min Small FFT, then go to a single iteration of 1 hour All FFT. If stable, it goes to overnight testing with one other core on the other CCD that needs testing (by default, Corecycler alternates between CCD1 and CCD2 so heat can dissipate). By the time I wake up, the two selected cores should be finished about 4-5 iterations each and the test log will tell all.
How the new profile looks so far, some rather interesting distributions especially on previously-unassuming Cores 2 and 6:
Is all this really necessary? I dunno, but because this is my main rig and not my sacrificial APU mem OC rig, I just don't tolerate unstable overclocks if I'm aware of them. Rounding errors are rounding errors. Lord knows Core 0 and Core 1 were quite unstable at -10/-10 over the past few months, that was very quickly obvious even in 6 min Small FFT.
The script also allows other tests such as AIDA and yCruncher. Being the pepega it is, I'll probably forgo AIDA, but yCruncher may be worth a looksie.
