So, I've been doing some fine tuning.
First I took that 3.9GHz fixed OC and turned it into a pstate OC. Changed P0 FID to 9B (3875MHz, easier to stabilize), left everything else as it is, disabled XFR and set a +0.1425v offset (base 1.1875v for the 1700) to reach ~1.31v under load taking vdroop into account, using LLC level 2. Voltage without load at that pstate is ~1.33-1.35v which is fine for a 24/7 OC. I'm not comfortable with more than that considering the long term, and 100.2*38.75 is around 3883MHz, and P95 stable. Also enabled relaxed EDC throttling, with enough cooling it's not a problem and increases performance.
CPU drops down to 1.5GHZ 1v at idle, perfect.
Up next, I did some memory fine tuning. BIOS defaults are ridiculously loose and there is a lot of performance left on the table.
This thread is amazing. Considering my memory kit is a 3200C14 factory OC'd to 3466C16, I found
The Stilt's recommended timings for B-die kits and applied them. 1.1v vSOC, 1.375v memory:
Ryzen loves fast memory and tight timings, performance at 3882MHz + agressive 3200C14 is better than 3900MHz with 3466C16 untuned defaults, as seen in CPUz, AIDA and cinebench.That's 1T without geardown mode, which is true 1T mode. BGSA is also enabled because I'm using 1DPC and both DIMMS are single rank. Latency is down 6-7ns, quite the improvement.
This is mostly stable, ran 16 instances of HCI memtest with 900MB each, made it to 450% coverage to have one error, the other 15 instances ran to 750% without problems then I closed them all. I need to keep stress testing.
I can't get 3200C14 to cold boot (PSU unplugged -> turn on) either setting primary timings to 14-14-14-30 1T leaving the rest auto or using the agressive timings. It will boot without problems if I don't unplug the PSU... there are some C6H BIOSes with cold boot fix, I guess that will make its way to the next official release. More AGESA revisions will continue to make things better... it's alright. I can always load the agressive profile after a cold boot and go on with my stuff.
The Stilt says P95 v28.10 with 128k FFTs in place is the proper P95 to use for Ryzen stress testing (CPU based at least) and it shows:
29.2 with official Ryzen support doesn't hit 150w. It seems the heatsink I'm using is just fine for this worst case, temperatures seem to be more than in check. VRM temperatures are fine, too.