More MH/s from Your GPU?

[Botanix]

I realized that many people make a mistake overclicking their cards not properly or setting up guiminers wrong. What I'm about to say was checked on 5870, 5850 and 5830. Every time there was a few % more mh/s when I used those settings.

Try this settings:

1. Core Clock - that's up to You.
2. Memory clock - 340MHz ( not less not more )
3. Extra flags: -v -w256
4. Assign all cards to one core.
5. No cross fire

5870,5850, 5830 seem to work as fast as they can on memory down clocked to 340Mhz - much faster. Anything below or abow is slower.

Examples:

XFX Hd 5870.

Core - 950
Memory - 340
Voltage - stock
Extra flags: -v -w256
CPU - Phenom x2 550 down clocked to 800MHz

GUIMiner - 433-438Mh/s each

Regards

[hmblm1245]

Have you OC them at all?

[Botanix]

Yes I have but without touching voltage.

2 XFX Hd 5870.

Core - 950
Memory - 340
Voltage - stock
Extra flags: -v -w256
CPU - Phenom x2 550 down clocked to 800MHz - 24W power consumption

Total consuption of my PC is 450W

around 870Mh/s

2x Sapphire HD 5850 Extreme

Core - 875
Memory - 340
Voltage - stock
Extra flags: -v -w256
CPU - Phenom x2 550 down clocked to 800MHz - 24W power consumption

Around 690MH/s

Total consuption of my PC is 320W

[jonand]

Any advice on fan settings?

Do you run the fans temperature controlled or at a steady speed?

Personally I have been running the fans at 70% steady speed the whole summer and kept temperatures decent on my 5850s. But I haven't been in that room... loud!

[Audriux9]

I am mining with few 5870, using MSI afterburner to overlock them. I tried to underlock memory clock to 340 to check performance, but unfortunately I was not able to set it, because MSI afterburner let me to underclock memory clock just to minimum at 600Mhz. Anyone else expierencing this issue?
Previously, I have used Sapphire trixx software, I managed to underclock my cards to 300Mhz with no problems, but with that software cards didn't work stable, and sometimes even locked up. MSI Afterburner has no issues with locking up, so I prefer to use it.

[Xephan]

Note that this only applies to dedicated mining rigs.

If you're using your multi-GPU rig for other purposes, then the memory clock on the GPU you're using for actual displays must be set higher or else you may experience flickering and other onscreen oddities. This is likely due to a UVD bug.

[Xephan]

I am mining with few 5870, using MSI afterburner to overlock them. I tried to underlock memory clock to 340 to check performance, but unfortunately I was not able to set it, because MSI afterburner let me to underclock memory clock just to minimum at 600Mhz. Anyone else expierencing this issue?
Previously, I have used Sapphire trixx software, I managed to underclock my cards to 300Mhz with no problems, but with that software cards didn't work stable, and sometimes even locked up. MSI Afterburner has no issues with locking up, so I prefer to use it.

Some cards just don't work well at certain speeds, my 5870 just goes crazy at 400Mhz but is ok at 390 and 418 for example. So you might want to try with Trixx but use say 310MHz instead. The exact clockspeed of the memory doesn't really matter, it's mainly to reduce the power wasted. 3xx was just a figure that came about as a nice 1/3 divider of a common 9xx ~10xx Mhz clock some people were originally clocking at.

[Audriux9]

I found something on that issue with setting memory clock. The cause of it probably is that I use Catalyst 11.7, and ass opposed here bitcointalk.org/index.php?topic=33611.0 not all multi GPU configurations work properly. So I'll wait for Catalyst 11.8, because even at the moment memory clocks are set at 600Mhz for 5870 and 500Mhz for 5970, I am satisfied with the temperatures and performance.


The question remains about the power consumption. I am looking for software to monitor it. And soon I'll try to play a bit with memory clocks.


Also I am curious, how to explain technically difference between extra flag parameter of -w 128 and -w 256, and impact to performance for 58xx and 5970 cards?

[Axez D. Nyde]

Also I am curious, how to explain technically difference between extra flag parameter of -w 128 and -w 256, and impact to performance for 58xx and 5970 cards?

I am, too! I guess it has to do with the way that things are scheduled on the stream processors?

Anyway, can u say something about the powersaving-effects of underclocking the memory?

[Mushoz]

Try Phoenix 1.50 with the phat switched and the improved kernel by Dia. Should yield even higher hashing rates!

[Xephan]

The question remains about the power consumption. I am looking for software to monitor it. And soon I'll try to play a bit with memory clocks.

This cannot be accurately monitored by software.

Also I am curious, how to explain technically difference between extra flag parameter of -w 128 and -w 256, and impact to performance for 58xx and 5970 cards?

Workgroup is an openCL parameter, very simplistically and likely not technically correctly, a workgroup of 128 works on 128 elements at once with all 128 able to share memory. So if the card can support larger workgroups, more elements can be operated on with the same data, hence faster in cases where this can be leveraged. Max workgroups is limited by the hardware, in steps of 32/64 it seems. And 256 seems to be the max in the 58/59 series.

[MiningBuddy]

I am mining with few 5870, using MSI afterburner to overlock them. I tried to underlock memory clock to 340 to check performance, but unfortunately I was not able to set it, because MSI afterburner let me to underclock memory clock just to minimum at 600Mhz. Anyone else expierencing this issue?
Previously, I have used Sapphire trixx software, I managed to underclock my cards to 300Mhz with no problems, but with that software cards didn't work stable, and sometimes even locked up. MSI Afterburner has no issues with locking up, so I prefer to use it.

See this guide on how to enable unofficial overclocking modes in MSI After Burner & bypass the limitations: https://bitcointalk.org/index.php?topic=18102.msg229511#msg229511

[Axez D. Nyde]

So if the card can support larger workgroups, more elements can be operated on with the same data, hence faster in cases where this can be leveraged.

Yet, and I am aware you did not explicitly say so, just want to point it out though, larger workgroups do not necessarily mean more Hash/s! Checked this on my card.

[Botanix]

"larger workgroups do not necessarily mean more Hash/s!"

That can be true but usually is not.

Try -w256 with 340MHz memory, that's the fastest and the most energy efficient for me.

"I am mining with few 5870, using MSI afterburner to overlock them. I tried to underlock memory clock to 340 to check performance, but unfortunately I was not able to set it, because MSI afterburner let me to underclock memory clock just to minimum at 600Mhz. Anyone else expierencing this issue?"

Try ACOLBF 1.75 You can download it here:

http://bitcointalk.org/index.php?topic=10264.0

Working for me. Can underclock down to 200MHz memory but still with workgroup 256bit, 340MHz is the fastest.

[Xephan]

So if the card can support larger workgroups, more elements can be operated on with the same data, hence faster in cases where this can be leveraged.

Yet, and I am aware you did not explicitly say so, just want to point it out though, larger workgroups do not necessarily mean more Hash/s! Checked this on my card.

That's why I had the "where it can be leveraged..." part

It also depends on other parameters. I would suspect on certain combination of memory/core speeds, a larger workgroup may take fractionally longer to load, which may then be amplified by latencies mismatches and processing time. It's like half a cycle delay means a full cycle wait on the next op kind of thing.

However on both my cards (5850,/5870), 256 is faster than 128. Which cards are you using at what clocks?

If your core clock isn't a multiple of your memory clock, try adjusting them to be. e.g. on my 5850 with 927/309 I get 377MH/s but with 927/340, I get 372 instead. Try it out for yourself, don't just trust numbers the rest of us put out because there are always differences in each of our systems.

[tarion3]

So if the card can support larger workgroups, more elements can be operated on with the same data, hence faster in cases where this can be leveraged.

Yet, and I am aware you did not explicitly say so, just want to point it out though, larger workgroups do not necessarily mean more Hash/s! Checked this on my card.

That's why I had the "where it can be leveraged..." part

It also depends on other parameters. I would suspect on certain combination of memory/core speeds, a larger workgroup may take fractionally longer to load, which may then be amplified by latencies mismatches and processing time. It's like half a cycle delay means a full cycle wait on the next op kind of thing.

However on both my cards (5850,/5870), 256 is faster than 128. Which cards are you using at what clocks?

If your core clock isn't a multiple of your memory clock, try adjusting them to be. e.g. on my 5850 with 927/309 I get 377MH/s but with 927/340, I get 372 instead. Try it out for yourself, don't just trust numbers the rest of us put out because there are always differences in each of our systems.

I really appreciate this suggestion of using a multiple of your memory clock. I've been trying to find a "sweet spot" for my two Sapphire 5830's for a while and didn't even think to do that. I've been running them at 935/300 (stock voltage) (getting around 303MH/s) and just switched it to 936/312. Once my mining pool comes back online, I'll test it out and keep tweaking. Thanks for the suggestion!

Along the lines of workgroup size, using phoenix 1.5 and phatk 2.2, the difference on either of my 5830's as far as Mh/s change when using either WORKGROUP=256 or WORKGROUP=128 is negligible to about 5 MH/s, with 256 being faster. I hear that 128 is faster with 68XX series cards, or cards with higher memory clock settings, but I don't have the cards to verify.

[Xephan]

Along the lines of workgroup size, using phoenix 1.5 and phatk 2.2, the difference on either of my 5830's as far as Mh/s change when using either WORKGROUP=256 or WORKGROUP=128 is negligible to about 5 MH/s, with 256 being faster. I hear that 128 is faster with 68XX series cards, or cards with higher memory clock settings, but I don't have the cards to verify.

If higher memory clock is the key, then we could test by changing memory clocks.

Switching to 900/1200 on my 5870
With 256 I get 391 MH/s
With 128 I get 398 MH/s

Dropping down to 900/900
with 256 I get 392 MH/s
with 128 I get 397 MH/s

Dropping back to 900/300
With 256 I get 405 MH/s
With 128 I get 399 MH/s

So it seems correct, higher memory clocks make w128 better than w256. But a really low memory clock gives better performance overall, maybe only for the 58xx cards.

[Axez D. Nyde]

Yeah I tested this too (5870) and most remarkably, the power comsumption of my system went down 50W.

I didn't spend too much time testing settings though, but I am running 920/340 now with lower power usage and cooler GPU temps. Did not touch the voltage though.

-k phatk DEVICE=0 VECTORS AGGRESSION=10 BFI_INT WORKSIZE=256

=> 419 MHash/s

Thanx alot for pointing this memory clockrate thing out, dear mister thread poster!

[Xephan]

Yeah I tested this too (5870) and most remarkably, the power comsumption of my system went down 50W.

I didn't spend too much time testing settings though, but I am running 920/340 now with lower power usage and cooler GPU temps. Did not touch the voltage though.

-k phatk DEVICE=0 VECTORS AGGRESSION=10 BFI_INT WORKSIZE=256

=> 419 MHash/s

Thanx alot for pointing this memory clockrate thing out, dear mister thread poster!

Could you try 921/307 and see what do you get?

[Axez D. Nyde]

I just realised I was lying earlier. I set the clocking back to 915/340 because the 920/240 setting kept freezing my OS.

anyway, tried your suggestion:
921/307 => 422 MHash/s
920/340 => 423 MHash/s

But no stable hashing possible.