5870 and 5850 Magical Clock speeds

[P4man]

A long time ago I saw a super nice graph plotting hashrate vs memory speed. I cant find it anymore, so I made my own. I made a little script to run cgminer with various core and memory settings and let it run until it had found 100 shares.

This is the result so far (ignore the right Y axis, left one is correct):








On hindsight 100 shares wasnt enough, but a few things are already clear;
- Differences are relatively mild. I havent been able to find ckolivas' 20 MH boost.
- its the ratio between gpu and memory clock that matters
- 300 and 150 are both lousy for a ~900-1000Mh 5870

Tested with:
- Linuxcoin (SDK 2.4)
- cgminer 2.2.1

[1714023668]

1714023668

[1714023668]

1714023668

[1714023668]

1714023668

[ssateneth]

you mean this graph? https://bitcointalk.org/index.php?topic=7964.0 first post

[P4man]

you mean this graph? https://bitcointalk.org/index.php?topic=7964.0 first post

Yep, thats the one I had in mind, thanks. Except his graph stops at 300,  while I tested everything under 300, so, phew, I didnt waste my time

No, how did he make that graph? He must have a more efficient way than me.

[QuantumFoam]

2.1 SDK?

edit: Nevermind, saw your footnote.

[deepceleron]

Although the highest performance is still at 295MHz RAM on my 5770 @ 980MHz, I just put it through some low MHz RAM benchmarking. One thing interesting is that you can't just judge the trend of a few sample points and interpolate between, you almost have to test each MHz setting one MHz at a time to not miss a peak. I found two more peak performance points below 295MHz, at 127MHz and 166MHz:

Mhz	Mhash/s
125	226.71
126	227.58
127	227.71
128	226.93
129	226.47
131	223.50
133	218.15
(295)	(228.28)

As you can see, if you missed testing that RAM peak by 5MHz, you might not even know it was there.

Catalyst 11.11/SDK2.5/Phoenix 1.7.4/phatk2 vectors aggression=12 worksize=256

[bangra]

When i start the 5870's in 1 of my rigs and just drop the memory drown to 300 i get a jump from 380mhash to 400 ish, Both cards are clocked at 868/1200 from factory, i see similar things when i do the same with the 5850's im running but going any lower always seems to cause stablity to drop on the cards

[rjk]

Cool, gained 100 mhash on a 6-card 5870 rig by dropping from 300 to 160.

[ssateneth]

Cool, gained 100 mhash on a 6-card 5870 rig by dropping from 300 to 160.

hmm, weird. I started looking for "magic" memory for my one 5870 that I wanted to mess around with in 1 mhz increments starting at 100mhz. I got up to 222 before seeing severe stability problems (crashes, miner errors, bsod) and stopped for tonight around 250. you can see my findings at https://docs.google.com/spreadsheet/ccc?key=0AjXdY6gpvmJ4dEo4OXhwdTlyeS1Vc1hDWV94akJHZFE&hl=en_US#gid=0

also have an initial graph here too if you dont want to look at numbers. I plan on doing vectors with worksizes of 64, 128, and 256 at low memory clocks, and vectors4 with the same worksizes at near-stock memory clocks. Any other situation is just entirely unoptimal with the 5870 in question. It may hold true for other VLIW5 hardware, probably different with VLIW4 hardware, and likely radically different with GCN hardware since I heard non-vectors is fastest with GCN.

Using Windows 7 x64, driver 12.1, 2.1 SDK, Phoenix 2.0.0 beta, phatk2 2.2 AGGRESSION=14, Radeon 5870 @ 1035 core

[deepceleron]

Cool, gained 100 mhash on a 6-card 5870 rig by dropping from 300 to 160.

hmm, weird. I started looking for "magic" memory for my one 5870 that I wanted to mess around with in 1 mhz increments starting at 100mhz. I got up to 222 before seeing severe stability problems (crashes, miner errors, bsod) and stopped for tonight around 250. you can see my findings at https://docs.google.com/spreadsheet/ccc?key=0AjXdY6gpvmJ4dEo4OXhwdTlyeS1Vc1hDWV94akJHZFE&hl=en_US#gid=0

also have an initial graph here too if you dont want to look at numbers.

Using Windows 7 x64, driver 12.1, 2.1 SDK, Phoenix 2.0.0 beta, phatk2 2.2 AGGRESSION=14, Radeon 5870 @ 1035 core

You are using 12.1, and likely it is using the 2.6 SDK DLLs that comes with that driver installed in the \windows directory and completely ignoring the 2.1 SDK that is just installed in Program Files and sits there. 1000MHz RAM will be where your only peak is with 2.6. I would quickly compare 300MHz against 1000MHz and make sure. 2.6 will be good at worksize 64 vectors4 @ 1000MHz and suck at 300MHz.

[QuantumFoam]

I've not had luck thus far on an xubuntu 11.04 install getting better than 430mh/s on a 5870 at 950. Downclocking below 300 on SDK 2.4 lowers hash rate, on SDK 2.1 it remains the same. I may mess around trying to get 2.1 reinstalled and give it another shot with the magic numbers mentioned in this thread. It may just be a linux thing, seems people on windows are getting much better mh on 5870s at 950.

[P4man]

I've not had luck thus far on an xubuntu 11.04 install getting better than 430mh/s on a 5870 at 950. Downclocking below 300 on SDK 2.4 lowers hash rate, on SDK 2.1 it remains the same. I may mess around trying to get 2.1 reinstalled and give it another shot with the magic numbers mentioned in this thread. It may just be a linux thing, seems people on windows are getting much better mh on 5870s at 950.

Above results were obtained on linux. FWIW, at 950/300 I got 433. 950/160 I got 439. SDK 2.1. 2.4

[bulanula]

I've not had luck thus far on an xubuntu 11.04 install getting better than 430mh/s on a 5870 at 950. Downclocking below 300 on SDK 2.4 lowers hash rate, on SDK 2.1 it remains the same. I may mess around trying to get 2.1 reinstalled and give it another shot with the magic numbers mentioned in this thread. It may just be a linux thing, seems people on windows are getting much better mh on 5870s at 950.

Above results were obtained on linux. FWIW, at 950/300 I got 433. 950/160 I got 439. SDK 2.1.

You seem to have forgotten that Linuxcoin comes with SDK 2.4 NOT 2.1 !

Cheers !

[P4man]

I've not had luck thus far on an xubuntu 11.04 install getting better than 430mh/s on a 5870 at 950. Downclocking below 300 on SDK 2.4 lowers hash rate, on SDK 2.1 it remains the same. I may mess around trying to get 2.1 reinstalled and give it another shot with the magic numbers mentioned in this thread. It may just be a linux thing, seems people on windows are getting much better mh on 5870s at 950.

Above results were obtained on linux. FWIW, at 950/300 I got 433. 950/160 I got 439. SDK 2.1.

You seem to have forgotten that Linuxcoin comes with SDK 2.4 NOT 2.1 !

Cheers !

I guess you are right. I just checked:
CL Platform version: OpenCL 1.1 AMD-APP-SDK-v2.4 (595.10)
Corrected in above posts.

[bulanula]

I've not had luck thus far on an xubuntu 11.04 install getting better than 430mh/s on a 5870 at 950. Downclocking below 300 on SDK 2.4 lowers hash rate, on SDK 2.1 it remains the same. I may mess around trying to get 2.1 reinstalled and give it another shot with the magic numbers mentioned in this thread. It may just be a linux thing, seems people on windows are getting much better mh on 5870s at 950.

Above results were obtained on linux. FWIW, at 950/300 I got 433. 950/160 I got 439. SDK 2.1.

You seem to have forgotten that Linuxcoin comes with SDK 2.4 NOT 2.1 !

Cheers !

I guess you are right. I just checked:
CL Platform version: OpenCL 1.1 AMD-APP-SDK-v2.4 (595.10)
Corrected in above posts.

Now that you corrected your mistake it would be FANTASTIC if you could test again using SDK 2.1 because that is what people say is better than 2.4.

Thank you !

[P4man]

Id like to see some evidence of that, in the form of a 5870 @ 1GHz performing over 462 MH/s sustained (or similar MHash/MHz) before Id even consider that.
Id also want a cgminer cypress .BIN file compiled with 2.1  SDK because backporting that on linuxcoin is going to be a pita.

If you can provide the above, I may give it a try.

[GenTarkin]

Anyone think this would affect a 5970?
I currently mine 810mhz / 190mhz and see about 371mh/s / GPU
SDK 2.5 , 11.12, vista 32, cgminer, phatk, w 256, v 2, intensity 9

[P4man]

Anyone think this would affect a 5970?
I currently mine 810mhz / 190mhz and see about 371mh/s / GPU
SDK 2.5 , 11.12, vista 32, cgminer, phatk, w 256, v 2, intensity 9

5970==2xunderclocked 5870
So yes.

That said, you already get 458 MHash per Gigahertz. Thats pretty much the same I got (~461), or close enough to ignore.
Unless SDK 2.1 would provide a non trivial boost, Id stick with what you have.

[GenTarkin]

Anyone think this would affect a 5970?
I currently mine 810mhz / 190mhz and see about 371mh/s / GPU
SDK 2.5 , 11.12, vista 32, cgminer, phatk, w 256, v 2, intensity 9

5970==2xunderclocked 5870
So yes.

That said, you already get 458 MHash per Gigahertz. Thats pretty much the same I got (~461), or close enough to ignore.
Unless SDK 2.1 would provide a non trivial boost, Id stick with what you have.

Cool thanks, I wont bother w/ older SDK cuz it has the CPU bug.