Dresdenboy
Golden Member
AGESA 1.0.0.6 is really good. I built my system last week and did some experiments. R7 1700X, Taichi (BIOS 2.36 now), 2x16 GB TridentZ 3200 CL15.AGESA 1.0.0.6 continues to bring interesting surprises.
After wasting a lot of time trying to get DDR4-3600 stable (not happening yet; I can get it to run almost everything except y-cruncher), I went back to the drawing board and decided to see what kind of subtimings AMD gave me using Ryzen Master to set DDR4-3200 14-14-14-32 (actually I went for 28 this time) as I had in the past. I went through my old routine, booted it up, confirmed that it was working, and then took a look in the UEFI to see what was going on.
Curiously the UEFI was showing the current timings/subtimings that had been set by Ryzen Master, but it was showing the DIMM's XMP settings as what was currently set in the UEFI. For example, it showed CAS/CL as 14 but the config field was set to 17. Booting into Windows asserted the Ryzen Master settings provided I changed nothing in the UEFI. Changing settings in the UEFI apparently "cancelled out" whatever Ryzen Master was trying to do. It did not overwrite settings changed by hand.
I can go back and record all those settings and post them here if you want to see exactly what Ryzen Master and the older pre-1.0.0.6 AGESA revisions were doing when setting the DDR4-3200 strap.
So anyway, I got 3200 stable (preditably) without changing anything. Using 1.4v vDIMM and 1.175v SoC voltage (LLC level 3) I got DDR4-3333 working with the exact same timings as what Ryzen Master set for DDR4-3200 14-14-14-28. Cool! AIDA64 reports ~52.8 GB/s read and 69.2 ns latency. Then I ran Firestrike to compare it to my Firestrike score from when I was running DDR4-3600 18-19-19-36 (1T and 2T; it made no difference at all).
The main thing that jumped out at me was the combined score. When running DDR4-3600, I only had the CPU at 3.6 GHz, and I got a combined score of over 5200 using a stock Powercolor 390. When running the CPU at 4 GHz with RAM at DDR4-3333 14-14-14-28, I got a combined score of ~4900 with the same Powercolor 390.
It seems to me that fabric speed is a major limiter of combined score performance in Firestrike.
That dual ranked RAM is running nicely at 3200 14-13-13-13-32-1T and 1.35V now, no VDD SOC increase was necessary. I should try 28, too.
Higher clocks didn't work so far. I tried ASRock settings (OC Tab), AMD CBS DRAM settings, and Ryzen Master.