So after a lot of fighting with various options with the X870E Pro Art Creator and 2x2x64 of F5-6400J3644F64GX2-TZ5RK here is what I found. Use case is quantization of models into quants that are optimal for gfx1201 via vllm, since no one seems to do that for the models I want to use.
This is currently all auto timings except primaries as I was hunting down the SOC voltage triggered errors. Nitro enabled, 8x,8x,guard band disabled to allow larger training range.
Aida bandwidth as is: ~67,000, 76,000, 72,000 92ms – essentially my baseline, from earlier testing I know there’s a ~80/92/75 73ns setup, but I had to backtrack to find the cause of the random error.
- Memory Context Restore leads to a failure to train/boot loop even with working timings.
- Proc ODT pull down 30 and pull up High-Z seems to work best.
- RTTs 5/5/3/5/4 seem to work best, the board actually corrected 5/5/2/5/5 into that despite manually setting them.
- High SOC voltage leads to hard to identify error, 1.28 = error in Kharau within firs 50%, 1.22 = still running without any other changes (CPU is well cooled, not a heat issue).
- G-skill heat blanket (their ‘heat sink’) removal = 5c reduced load/idle temps. Use a heat gun and slowly release the adhesive to remove the stupid heat blankets. Fan pointed at ram has more efficacy without the heat blankets due to direct airflow exposure. ~32c idle / ~41c 2 hours into Kharu testing with a 120mm 1200 RPM fan pointed at them.
- Asus will adjust the trfc value to == 409.xx, just let it for stability until all other settings are done being tuned.
- The Samsung M die seem to tolerate the max 65536 trefirefresh interval just fine at 1.35vdd/vddq or higher.
- Tcl does scale with voltage, but not as well as Hynix A/M 16/24 chips do.
- TRTP is very picky with these sticks, does post at 20, still testing. In general, secondary/tertiaries have very little wiggle room, more than the terrible Micron 2x64 kits, but still not much.