Nice, happy you found a version that worked for you!

Definitely. Flash your Vega 56 with 64 air bios.
It'll help as it will increase the memory frequency of the card.
In 56 bios, I was able to hit 935 MHz but with 64 air bios flashed the memory frequency was stable at 1100MHz.
Any further and it was unstable.

Ok - one other possibility...  somewhere between the TRM implementation, and newer versions of drivers, I feel like I’ve seen instances of transitional states lighting up long enough for voltage settings to actually matter.  It’s possible you’re crashing during the p4-p5 transition going from 800 to 1107, as 820mv may be too low (or something like that.) 

Try setting either a more graduated voltage list, or set every state to your final voltage (I would suggest the former, in keeping w/ the PPT as floor concept.)

What about when you run every state @ 900?  What does hwinfo say about SOC in that case?  Or are you sayin opening hwinfo while mining causes a crash?  If so, try running hwinfo first

I’m betting the issue w/ being stuck @ 900mv is due to you having set p0 @ 900.  Even though you have lower states locked out in ODNT, you would already be at 900 in your idle state, before ODNT is ever involved.  This may very well prevent you from ever going lower.

Best practice is to have voltages in PPT be coded as a floor - this allows it to serve its purpose (removing the uv barrier,) while staying out of the way otherwise.  Your idea of setting 820 across the board except for top states is moving in the right direction (though I believe default p0 is actually 800, so 820 is still high.). As for why that approach causes crashes - only thing I can think is that 900 has been a threshold voltage for 1200 SOC in my exp.  Are you sure you’re running 1107 SOC?  As in, that’s what’s reported in hwinfo under load?

Miner vs poolside testing is a b*tch, and I dedicated the last section of the CN_MAX_YOUR_VEGA.txt guide in the 0.4.4 release to it. That said, your numbers are way off, this should happen in <= 0.07% of all cases _if you were running at a static diff_. Since you aren't, it's really difficult to reason about if this is a "normal outlier" that happens every 100th run, or that it more or less proves that something clearly is off.

Drilling deeper and looking at each gpu, two of your cards look a little fishy at 1480 h/s poolside, but we have way too little data to assert anything on a per-gpu basis. So, please run a xmrig-proxy test as decribed in the guide for e.g. 50k shares total and you should have the data you need - it might be that those two cards are not doing great at the current clocks and/or timing mods, and it should be clearly visible with that higher amount of shares. We're doing tons and tons of testing ourselves to make sure we hit the poolside hashrate, I'd love to get more user data here, but it must be from runs that I know I can trust from a statistical standpoint.


Newer drivers often seem to lock P7 core+P3 mem during load to not be too aggressive with switching p-states.

For the PPT itself, this is an artform. Don't be fooled by the view in ODNT. The only thing that exists is a single vdd lookup table which is _shared_ by core and mem states. ODNT fools you into believing that you can separate voltage settings for e.g. core P4 from mem P3, which isn't true unless you do some serious PPT tweaking. I'm not 100% it's possible even then.

To simplify your tests, I'd either (1) use the same target voltage for ALL core and mem states in your PPT or (2) make sure all core and mem states have increasing voltage levels in respective series and also that core P4 voltage == mem P3 voltage and that it matches your target.

I'm not 100% sure what happens in your tests. Maybe mem P3 is rather grabbed from core P4, so in your first test you get 900 mV (~= 895 mV), and in your second test you get 820 mV which is way too low to handle sustained pressure?

That’s an insane voltage range, and I can’t fathom how 970mv could possibly be necessary. Between 9 64s and a flashed 56, I run avg 834mv, none over 850, for similar clocks/timings.  That puts me @ 178w per GPU at the wall.

At 970, you’re running at least 35% higher power, meaning over 240w per card.

It sounds like you have some ppt issues or something, maybe you’re running your SOC @ 1200mhz unnecessarily (it should be 1107 @ your clocks.)  Although, even then, you prob shouldn’t need much more than 925mv.

I recommend to go online and search for a simple tutorial on how to install amdgpu drivers. It's not that hard but i can imagine if it's your first time, that it might be a bit tricky.

I'm working on something to make this process easier.  ;-)



Hey Mashy, sorry sir, no XMR address.
You can always use a shift service if you want to tip.
coinswap, changelly.. etc..



Who knows.. perhaps something is coming  

what gpu you have? vega 56 samssung or hynix?

start WinAMDTweak.exe --i 0 --rp 10 --rc 44 --rfc 300 --rrd 3 --rcdrd 12 --rcdwr 5
and
start WinAMDTweak.exe --i 1 --rp 12 --rc 44 --rfc 300 --rrd 3 --rcdrd 12 --rcdwr 5
and start
WinAMDTweak.exe --i 2,3 --rp 14 --rc 44 --rfc 300 --rrd 3 --rcdrd 12 --rcdwr 5