@wendell There is one thing that you could help me with, because I am stumped for an explanation.
ScatterBencher made a video on overclocking the 9700X and so I decided to see what I could get out my 9950X if I disabled one CCD.
Now ScatterBencher was doing the whole PBO + CO thing whereas I was configuring my 9950X my way, which admittedly gives a tad less Single Core performance, but a lot more Multi Core performance.
Be that as it may, ScatterBencher ran PyPrime which is a Single Core workload that runs at RealTime priority and his CPU should have been boosting to about 5.7 GHz.
My CPU was confined to a max single core clockspeed of 5.55 GHz.
So here was his result (lower is better) with his stock result in blue and his overclocked result in green:
It gets even weirder though. When I went back to 16 Cores/32 Threads with my 9950X where it runs at a max clockspeed of 5.375 GHz I got the following result:
I’m on AGESA 1.2.0.1a, which now enables me to configure my 9950X completely in the BIOS.
This was not the case with AGESA 1.2.0.0a Patch A.
WIth regard to the latency, I have no idea, because I don’t have access to the tool Anandtech was using.
So it is a case of “Trust me bro” on the part of Anandtech, and considering the Tech Media’s track record when it comes to Ryzen over the past five years - and calling it pathetic would be a compliment - I wouldn’t put much stock in the results; especially considering that they were running on WinTel 11.
Try 6200/2066. It might have better latencies since FCLK is matched 2:3 to the memory speed.
6200 MT/s UCLK=MEMCLK FCLK 2167
This one does not make sense to me. It should not be so slow. Probably it is unstable. On zen 4 (and I assume zen 5) infinity fabric has some error correction mechanisms that slow down the CPU.
Possibly you can gain some performance in the secondary/tertiary timings but tuning those can be very tedious. You could try the infamous buildzoid timings
or this sheet with suggested/calculated timings
But the most important timing is tREFI and tRFC which you have already tuned.
If 6200 is stable at VSOC 1.2, and 6400 UCLK=MCLK boots, you could probably get 6400 stable with somewhat higher VSOC. Just hypothetically of course
Pyprime is single-threaded AFAIK so higher single-core might help a bit, but it is mostly a memory latency benchmark. y-cruncher is more bandwidth sensitive, if you want to cover that too.
