Optimizing GPUs for maximum hashrate

[gsrcrxsi]

This is what I did.

1. Clean up the formatting in the AverageHashrate files by removing the line spacing. Make sure the spacing of each field lines up in straight vertical columns. Save the file.
2. In Excel, import the text file. Data tab -> import .txt
3. Navigate to the file when Excel prompts you and open it.
4. Import using the default setting "fixed width" I think is the term. NOT the comma deliminated.
5. Move the bars and isolate the fields for core clock (or mem or TC, whichever you're working with) and the (avg) Hashrate. Make sure that you don't get any colons or letters in there.
6. Click finish to import the data.
7. Delete the columns of "extra" data leaving only the columns of core (or mem or TC), and hashrates.
8. Create a scatter plot and plot clocks on the X and hashrate on the Y

 

[gsrcrxsi]

so to get long term stability, ive had to make some tweaks. getting the best numbers is one thing. keeping it stable is entirely different.

MSI 7870 - stability not limited by temps. with the fan at 85%, it only gets to about 55C. I had to reduce the intensity from 20 to 19. at 20, it was crashing the driver every few hours. at 19, it has been going 24+ without a hitch, only lost 5kh/s by dropping it. Since the card is also underclocked to 890MHz, ive slightly undervolted it from 1.25v -> 1.213v. I may try to drop it down even further. but im not sure how the card handles memory voltage (if its a multiplier of core voltage or something), so i'll have to be careful since its overclocked. 383kh/s

XFX 7950 - no tweaks, still running 1120/1480, 1.091v, -I20, 665kh/s.

ATI 5970 - this card im still working on. but its been a bit of a nightmare. the max settings of 815/1000 -I20 just cant be kept stable. this card has a stock voltage of 1.05v on each GPU. which doesnt seem to be enough to support the overclock. ive found a "better" voltage range around 1.10-1.175v area, but it will still crash the driver after 1-2hrs. sometimes less. and the VRMs are getting VERY hot. like 100+C hot. I'm not positive, but im going to guess that the AC heatsink i have on there just isnt cooling the VRM's properly, despite that it's cooling the cores VERY well. the cores cant even break 60C, but the VRMs are getting crazy hot. this greatly reduces overall efficiency and increases the power consumption. so far, ive got the settings down to 800/1000 and 1.175v, -I19. and it'll go a few hours.

the biggest problem is that when the 5970 crashes, it doesnt recover to the settings i want. what happens is that GPU1 usage will drop to 0, then the driver crashes and GPU2 will drop to zero but resets the core to 725/1000/1.050 (stock). leaving GPU1 at the old settings. then cgwatcher sees the problem, and resets it, but cgminer wont, or cant, change the settings back, even though it's trying to. the only thing that fixes it is a hard reboot. in which case i can then set the clocks and voltage to whatever i want. voltage is set by MSI afterburner via a saved profile set to apply at startup, clocks are set by cgminer when it starts up. cgwatcher is set to reboot the computer if it detects the total hashrate drops below normal (which indicates that the core clock was reset to 725). i only loose ~30 seconds for the reboot.

i really wish there was a way to have cgwatcher reboot the computer after x number of hours. but i guess i cant complain for a free card. its also entirely possible that the card is just damaged and hanging on by a thread. i DID have to bake it back to life after all.


all in all, im happy with the results. so far ive been able to squeeze an extra 90-100kh/s between all of the cards. bringing me from 1.7 -> 1.8MH/s

 

[xsuperbgx]

For stability on windows miners I found that going into power control settings and set turn off monitor to never seemed to greatly improve my stability.

 

[gsrcrxsi]

i already do that.

 

[Janus67]

So I ran the test on one of my 290X cards and the graphs came back with 860 core being the way to go (now this was with 24000TC [the default it had] when I normally use 33792) and had lower amounts going up from there. Tried it and my khs was terrible (100 lower than the same settings with the core I normally use). Not sure what would have gone wrong, I followed the instructions about using a safe memory (1300), set powertune to 20, and didn't change anything else from the settings that it had in it.

Edit: just remembered that it auto-defaulted to worksize of 256 when I normally use 512, although it still didn't make a difference when I tested those settings. /shrug, just hard to waste a day of slightly less efficient mining to figure it out.

 

[xsuperbgx]

I found that you have to rerun each test until you have everything dialed in. Set everything as you have it normally and run each test to dial it in. Every time you change one setting it seems to throw the others off. So set your tc and best mem speed,then run for GPU. Set that then run and see if there is a better mem., last run tc and see if anything is better. I never finished the process as I felt I wasn't getting anything. I may mess with it again if I get some time.

 

[gsrcrxsi]

i usually use whatever TC i had been previously using.

but the TRENDS seem to translate for really any setting. look at my 7870 plots. it showed that across 2 separate and wildly different settings, that core clocks in the 850-900 range where where i should be.

some fine tuning, if only for stability, should be expected.

 

[gsrcrxsi]

Fruits of my labor so far:



you can see that theres a LOT more stability since last night. and much less noise in the share submission. this will translate into slightly better performance and slightly better profits. and most importantly, hopefully a lot less time that i have to deal with it. having the rig go down hard while you're out of town really sucks.

 

[Silver_Pharaoh]

Fruits of my labor so far:



you can see that theres a LOT more stability since last night. and much less noise in the share submission. this will translate into slightly better performance and slightly better profits. and most importantly, hopefully a lot less time that i have to deal with it. having the rig go down hard while you're out of town really sucks.

WOW!
The fruit looks tasty!

That's some nice btc right there

 

[ZL1]

gsrcrxsi, what was your end result for the 7870 ? because with 950/1250, i19, tc15232, g1, w256 mine pulls 400khs steady and thats a guess+word of mouth settings

 

[gsrcrxsi]

380kh

890/1370, i19, g1, TC 24000.

It did 390 with i20 but wasn't stable. Would crash a lot. Been running 3 days straight with no issues at 19 though.

Maybe I'll play with it more, but according to my tests , 870-890 was the sweet spot.

 

[ZL1]

can you for the sake of it try my settings ? see if it pushes 400 then, bc like i said before, i worry about the way this soft looks for the sweet spot and i feel it doesnt find the exact one, better than stock of course, but not the sweetest spot

 

[gsrcrxsi]

On another note:

After running TC tests on 3 different cards, I'm fairly confident that it's really not that important. Sure some values are better than others, but there no sweet spot like core/mem. The most important thing with TC is that it's high enough to keep you from getting HW errors. For example, on my 7870, I saw absolutely no difference between 15232 and 24000. You can probably just pick a decently high value and play around within 100-200 and hone in on a value that will be fine.

 

[gsrcrxsi]

can you for the sake of it try my settings ? see if it pushes 400 then, bc like i said before, i worry about the way this soft looks for the sweet spot and i feel it doesnt find the exact one, better than stock of course, but not the sweetest spot

Sure I'll try it out when I get home

 

[bob4933]

I don't know crap about mining, but why would you run your card 28% harder for (literally) less than 1% gain?

 

[gsrcrxsi]

Stock clocks are 1100MHz. It's already underworked anyway.

 

[gsrcrxsi]

Also, from 380->400 is a 5.2% increase.

 

[ivanlabrie]

The Stilt (a famed ocer and tweaking guru of sorts) posted a memory latency optimization thread at litecointalk, and he debunked the ram/core ratio sweetspot theory.
He says the problem is badly calibrated auto mem timings when changing straps, so different frequencies will impact latency for the worse, despite going up on ram clocks.

He did modify a lot of 280X and 7950 bioses, and he'll soon began working on R9 290/x cards. I subscribe to his line of thinking, he's shown great success with his method so far.

Then there are optimized scrypt.cl files for sgminer, those seem to help. I've seen plenty of 7950s pulling 710kh/s and a few r9 280x cards doing close to 800kh/s and slightly more with that and Stilt's bioses.

 

[gsrcrxsi]

i dont think theres a core/mem "sweetspot" in the sense that theres a specific ratio or something like that.

but as my tests have been showing, there certainly are areas on both the core and mem that perform better. look at my 7870 core graph. it performs better when underclocked, and that seems to be the trend at ALL memory speeds. thats why it's worked for me.

who is stilt? never heard of him. does he charge for BIOSes? or are they free? i've also been curious about these sgminer optimizations, but i kinda dont want to rock the boat lol.

ZL1:

i tried 2 settings:

950/1370, i19, tc 15232, g1 ---- 380kh/s (you can see i left my memory the same)
950/1250, i19, tc 15232, g1 ---- 310kh/s, this performed worse.

 

[Silver_Pharaoh]

There is no sweet spot.

At least, there isn't one on my 6850
I downclocked the memory to 900 Mhz, for some energy savings....

Okay, what's the equivilant to a 270x? A 7850?
I just want something cheap with 400+ Kh....