pointer to understanding GPU power optimizing?

greybeardgreybeard Member Posts: 18
Greetings,

For an old man, I think I made pretty good progress in the past week. Sorry if I ask stupid questions. Working with (4) Gigabyte RX580 Gaming 8GB that have Hynix memory. With geth/ethminer in linux I was getting ~20mh/s each GPU, stock vbios.

Reading here, I learned how to flash vbios. I used latest atiWinflash for saving factory roms, and latest polaris bios editor for writing new roms. Found some settings here and anorak that helped boost hash rates. Windows drivers also helped. Stock vbios in the cards on Win10, now I get ~23 mh/s each. Upgrade to AMD Blockchain drivers for Win10, now each card 25 mh/s.

Flash vbios with some magic "timing straps" whatever the hell that is, and now I get 28.3 mh/s each card.

One out of four cards, I made a brick but learned to unbrick it , and back in bizness. :-) To all persons who share knowledge here, thank you. <3 Saved my ass! :-D

Now stupid questions about reducing power yet maintaining hash rate. I get that each card is different. I kinda get how the bios changes can improve power consumption, but not really.. Using gpu-z tool, gpu power consumption figures when mining look like this; (one card at a time for bios tweaking and testing).

card0 130W
card1 128W
card2 126W
card3 126W

Now that all cards are flashed with custom timing straps, hash rates are within 0.5 between cards. But power draw obviously is not. I am very happy to have gained 8mh/s per card by drivers and vbios flashing. But I don't really know what I'm doing, and don't have experience with mining. Is it normal to have that much power variance between cards? Starting points for tweaking: GPU core clock: 1350 Mem clock: 2200 undervolt mem to 950 - all these done in vbios that is flashed onto card.

Sorry I cannot seem to use AMD Settings "WattMan" for the undervolting, the app does not seem to recognize mouse click or I just don't get how to change settings using that software..

I have not messed with claymore volts features yet. I *have* messed with claymore fan control, and it seems to work ok. No over heat probs so far, temps ~65C.

Any suggestions on reducing power without hurting mh/s are super appreciated. Thanks again to all who share how to do this crazy stuff.

Comments

  • ciprianptciprianpt Member Posts: 214 ✭✭
    hey @greybeard , the main reason for your power consumption is the core clock, 1350 is too high and makes your cards to run hot. You should lower it to 1150 or 1130.

    Are you mining ethereum only ?

    try adding this line in claymore's .bat file -cclock 1130 -mclock 2200 -cvddc 850 -mvddc 875

    Let me know your results!

  • greybeardgreybeard Member Posts: 18
    Hey @ciprianpt, thanks for the clue.

    I dont know why but tweaking any of those volts/clocks parameters in claymore actually reduces hashrate in my setup.

    So I came at it the other way. Figuring out how to use WattMan.

    I got 2 cards to come down in power, but with same settings, 2 cards went up in power.. I must be patient and go one step at a time.

    Oh yeah, and with this round of power tweak, all hashrates now hit 31 sometimes. They don't seem to sustain that rate. More tweaking is in order. Thanks again for the pointer, much appreciated.
  • greybeardgreybeard Member Posts: 18
    after fussing with WattMan, I got all cards between 100-111W according to gpu-z, and hashrates > 30 each card. Each card has slightly different gpu negative freqency and power reduction. But I'm happy. kill-a-watt arrives today maybe I'll see what the rig draws at the wall.
  • ciprianptciprianpt Member Posts: 214 ✭✭
    greybeard said:

    after fussing with WattMan, I got all cards between 100-111W according to gpu-z, and hashrates > 30 each card. Each card has slightly different gpu negative freqency and power reduction. But I'm happy. kill-a-watt arrives today maybe I'll see what the rig draws at the wall.

    Awesome :) if you want me to take a look message me your teamviewer details..
  • greybeardgreybeard Member Posts: 18
    These gpu are touchy, or the Wattman software is kinda sketch, or both. In any case, I got all 4 cards down to ~103W each :-) and kept hash rates just above 30. With 4 gpu, total system at the wall is 675W. My guess, with 8 gpu, total system watts will be just under 1200. Does it seem reasonable?

    I can get the cards to 100W each but Claymore got hung up on gpu0 (my lowest power card) and I had to reboot. So I gave that one a bit more gpu freq, a bit less mem freq. Now they are all balanced ~105W, and spread between cards is < 2W, and within 0.5 mh/s of each other. But wattman is fussy shit.

    Also, if I'm not mistaken, there was recently a new "mining round" and it got more difficult; is it true? If yes, it would explain the slight decrease in mh/sec today.
  • ciprianptciprianpt Member Posts: 214 ✭✭
    For gpu0 you can just change memory undervolt from 875 to 900 if it hangs again do the same for core clock by increasing it to 875 like:

    -cclock 1130 -mclock 2200 -cvddc 875,850,850,850,850,850 -mvddc 900,875,875,875,875,875
  • headshot155headshot155 Member Posts: 76
    Your wattage of 675 at the wall means your cards are probably drawing 140watts each. A motherboard will typically consumer 100 watts so the calculation is 675-100 / 4. This large discrepancy between reported watts in gpuz and what’s at the wall is because your risers consume a fair bit of power too. So I reckon a 8 card system would be 1250, so not far off your estimates
  • ecs87ecs87 Dekalb, ILMember Posts: 273 ✭✭
    ciprianpt said:

    For gpu0 you can just change memory undervolt from 875 to 900 if it hangs again do the same for core clock by increasing it to 875 like:

    -cclock 1130 -mclock 2200 -cvddc 875,850,850,850,850,850 -mvddc 900,875,875,875,875,875

    Does anyone else see issue with this configuration?...or am I just getting old?

    As far as I know, the drivers default to highest voltage value between core and memory (for Polaris cards at least) for each power level if you're setting voltage through the ADL. Therefore your cvddc values are doing absolutely nothing since the mvddc values are 25mv higher on each card (which begs the point: why even set both? just set one and done)
  • greybeardgreybeard Member Posts: 18
    @headshot155 , thanks for that calculation advice. Good to hear that my estimate is rational.

    I like to measure the power externally if possible. So I made a gpu power meter that can measure (1) pos/neg pair feeding a gpu. Evga sent many fine cables with the PSU, so I altered one. ;-)


    Seeing ~20V at 7A on each pair of pos/neg wires when mining, so.. ~84W power going to the 8pin connector on top of gpu. And for each gpu to be drawing ~140W total, each card must also be pulling ~60W from the mobo slot and riser. Cool. I feel better understanding that.

    @ciprianpt , for whatever reason, my setup does not tolerate making changes to gpu clock values using claymore. Hashrates just get worse, immediately. I am able to change fan speed using claymore with no negative effects, +/- 5% accuracy.
  • KlintistwoodKlintistwood Member Posts: 130
    @greybeard I'm using a separated PSU for my GPUs and thanks to this setup I can measure the exact consumption of all GPU directly into the wall. That spares me some trouble finding creative solutions like yours :)
  • rmhrmh Member Posts: 223 ✭✭
    greybeard said:


    I like to measure the power externally if possible. So I made a gpu power meter that can measure (1) pos/neg pair feeding a gpu. Evga sent many fine cables with the PSU, so I altered one. ;-)

    Your measuring stuff is very cool! Congrats!
  • ciprianptciprianpt Member Posts: 214 ✭✭
    @greybeard can you build one for me ship it? Ill pay for it of course :) thanks!
  • greybeardgreybeard Member Posts: 18
    @ciprianpt , you don't need me to build one - you can do it yourself! It's not that hard. I made this device to avoid using a multi-meter, because I cannot see that well anymore, and my hands shake a bit.

    You can achieve same measuring results without hacking a cable simply by using a multi-meter. I can help with that if needed.
Sign In or Register to comment.