Optimizing GPUs for maximum hashrate

[gsrcrxsi]

Very interresting, what Volts are running, did i miss that or did you not mention it?

Since i have 2x 7950's i need to try this when I have time.

in step 1 i reflashed the BIOS to one that i modded with VBE7.

running 1.09v

 

[gsrcrxsi]

I did, but for some reason, the core got stuck at 930 Mhz

There was no difference with TC's any higher than the 6144 setting I use...

it was probably unstable at those speeds.

 

[gsrcrxsi]

uploaded the TC plot for the 5970, same story, doesnt much matter with the exception of a few areas.

 

[ZL1]

ok, I cant seem to get around to trying this yet, so I will just ask you
I gather this is very similar to those older auto-oc tools, which bumped core til it held no more, then bumped ram til it held no more, does it do the same thing ?? first one then the other ? or does it try ratios and mixes ?
heres why I ask, take the 5830, its a great lil card really, but its peak perf is achieved by overclocking core to 890 and downclocking ram to 890, if you were to only bump the core, then the khs would be lower

 

[Silver_Pharaoh]

ok, I cant seem to get around to trying this yet, so I will just ask you
I gather this is very similar to those older auto-oc tools, which bumped core til it held no more, then bumped ram til it held no more, does it do the same thing ?? first one then the other ? or does it try ratios and mixes ?
heres why I ask, take the 5830, its a great lil card really, but its peak perf is achieved by overclocking core to 890 and downclocking ram to 890, if you were to only bump the core, then the khs would be lower

Pretty much.

OC's the core, then the mem, then finds the best TC.

Although, I've noticed memory speed has no impact at all on hashrates.... At least, on my 6850. My memory is downclocked to a mere 550Mhz

 

[ZL1]

Pretty much.

OC's the core, then the mem, then finds the best TC.

Although, I've noticed memory speed has no impact at all on hashrates.... At least, on my 6850. My memory is downclocked to a mere 550Mhz

thats bad then, because the optimal speed is usually inbetween min and max OC for both ..
it would need to try combinations to a near insane number, like start from 600core and 600ram and keep trying 605core 600 ram, then 600 core 605 ram, then 605 core, 605 ram, and so on; only trying 605 core with 605 ram isnt good, it may be highest stable oc, but its not the best mining in many, many, many cases

 

[gsrcrxsi]

You can set the start and stop values to whatever you like. I purposely set my start clocks very low (200MHz underclock).

I'm inclined to believe that the cards that see the performance benefit from weird combinations still follow the pattern given to you. So you can see in my 7950 memory plot. There's a nice little peak at 1250, then it dumps, and slowly climbs to another peak at 1480. I'm willing to bet you'll see the same pattern regardless of the core clock chosen.

Give it a shot.
Run your core clock (with memory at stock). It will prob show max output at around 890.

It's not about max OC. It's finding the peak KH/s output per speed. Look at my 7950 core clock. It shows just that. I can OC it past 1120 just fine. But the Hashrate drops.

 

[ZL1]

ya, but if it hits max and stops there and the just plays ram, it wont know it needs to downclock alongway
but ya, I gotta just do it and see how it does, the 5830 is perfect for this, I know its best speed, which is the perfect weird ratio of 1:1 (gpu oc, ram uc)
i will post once i do it, lets hope the soft figures it out )

 

[gsrcrxsi]

You run them manually. It's 3 different batch files and you have to manually set the parameters on each one.

It's not automatic in the sense that you're thinking.

 

[gsrcrxsi]

Also. Maybe you're confused by my plots.

I made those myself with Excel lol. They aren't created by this tool. All the tool does is test all the settings and spit the results into a text file. It's you're job to read the text files and find the best results from each test. The tool doesn't pick any "best" value for you and move on or anything like that.

I just made the plots because it's interesting and easy to interpret.

 

[gsrcrxsi]

aaaaand testing of the 7870 begins.

this will be a special test. in that i will test it with 2 major configs. while setting this card up, i got the best performance running a -I12/-g2 setup rather than the more typical -I20/-g1 setup (which i did try at first).

so i'm going to run through the tests at -I12/-g2, and then at -I20/-g1 and see what provides the best results.

 

[Silver_Pharaoh]

Ummm I hate to hijack, but is the powertune setting supported by the 6xxx models?

Cgminer says "failed to modify powertune settings"???

 

[gsrcrxsi]

im not sure. do you use powertune at all in your personal setup now? whatever you do change the default 50 value in the bat file to 20. 50 caused all kinds of problems for me.

if you want to remove the powertune setting, just open up the miner bat file and remove that argument.

 

[gsrcrxsi]

Looks like I'm getting closer to stability on the 5970. This is a tricky card.

I've been running stock voltage 1.050, it's unlocked. Guess I shouldn't have expected to be able to OC a dual GPU card without bumping the voltage a bit. It seems to be a very delicate balance though. Too low and it's unstable. Too high and the VRMs overheat (100C) and cause instability of their own :/.

So far 1.112v seems ok. VRMs "only" 90C now.

 

[Silver_Pharaoh]

im not sure. do you use powertune at all in your personal setup now? whatever you do change the default 50 value in the bat file to 20. 50 caused all kinds of problems for me.

if you want to remove the powertune setting, just open up the miner bat file and remove that argument.

I never ran the powertune setting till I tried it yesterday and got that error

 

[gsrcrxsi]

ok, some preliminary results from the 7870. ran both the clock and mem tests with -I12, but the results werent what i expected. i was expecting something similar to before, but i got results that indicated that almost any setting didnt make much difference.

i have theorized that -I12 may be a bit too low, so i bumped it to -I13 and will check the results. if that fails, then i will just move on to the -g1 tests.

so far:

 

[Silver_Pharaoh]

Uhhhh, It looks like the best hashrate was around 1100 Mhz??

 

[gsrcrxsi]

It's a close call.

It was actually a 2-3 KH higher where I've indicated.

 

[gungeek]

For the 7870, you may want to start sweeping at 900 instead of 1000. I've found a sweet spot at 955 on 2 sapphire 7870s. Both are 400 Kh/s at 955/1250 i-19 g-1 tc-15232. It allows good performance and you can undervolt to reduce heat/temps.

 

[gsrcrxsi]

To be clear. Are you saying to sweep 900-1000 for the memory or the core clock?