Author Topic: xfx cards  (Read 1389 times)

Offline fuzznarf

  • Jr. Member
  • **
  • Posts: 52
xfx cards
« on: April 11, 2013, 02:07:04 PM »
So i picked up an XFX 7970 yesterday to go with my Gigabyte 7970 in my big rig.  The weird thing is that using the same clock settings and the exact same batch file to run it I am getting about 300khs LESS than my gigabyte.  Has anyone seen this problem?

I tried a bunch of different settings and the best I could do with the card was about 450KHs. 

as a side note, I have some 7850 XFX cards that run at 300KHs on stock clocks (even though they run hot 85ยบ+), which is about what I would expect, but this 7970 really has me stumped... any help would be greatly appreciated. Thanks!
LTC: LfvVoS9zGeRqXDyKpsvSQNXoyoNBgEeQc6
BTC: 1Bj1xQjQg8eJqZNkuozNEwLN58FYKqzxsz

Offline gfoot

  • Jr. Member
  • **
  • Posts: 165
Re: xfx cards
« Reply #1 on: April 11, 2013, 03:11:04 PM »
I have one too, it has been quite troublesome.  I would be interested to know whether yours acts similarly to mine - I kind of wondered whether the idiosyncracy was due to my other hardware, which is all about five years old - 650w power supply, Phenom X3 CPU, 4GB DDR2 RAM, average motherboard.  But if it's just the way these cards are then it's disappointing and I wouldn't recommend them.  I picked it because it looked capable of keeping itself cool; I was put off the cheaper MSI (which sounds much more miner-friendly in retrospect) by its single snail fan design.

Firstly, make sure you set the powertune to +20%, that should get you to about 550KH/s.  Then try low thread concurrency (e.g. 3072) and high -g value (e.g. 9), it should get to at least 600KH/s.  You can try lower -g values with higher thread concurrencies, but it seemed to me that reducing the thread concurrency has very little effect on the hash rate, so is not a bad trade-off in exchange for the positive effect of a higher -g value.  I believe the limitation is the product of the two, i.e. 9 * 3072 = about 32k, perhaps some memory limit kicks in at that point.  In particular, cgminer's default mechanism for setting thread concurrency doesn't seem to account for -g, so you have to set them both by hand.

As usual for 7970s, -I13 seems like the sweet spot, regardless of thread concurrency (although I think ckolivas said they were related somewhere).  Values like -I14 and -I12 were consistently much worse than -I13, but in some situations I did find that -I18 or thereabouts gave similar performance to -I13.

Slight memory overclocking destroys performance (e.g. 1501 instead of 1500 loses about 20% of the hash rate!), but larger amounts of overclock (1600, 1700, 1800) seem to get the performance back again.  Even so I never saw a significant benefit from overclocking either the memory or the core.  Core clocks up to 1160 were stable for me, 1170 would swiftly hang the GPU.  I tried adjusting it over a wide range in +5 or +1 increments but overall compared to the default 1050 I only observed maybe 10KH/s improvement through that.  So I'm running memory at 1500 and core at 1150.

I also found worksize 256 gave some benefit; the default of 64 was OK though.  128 was terrible.

I found on my card the voltage is also fixed, so no way to reduce power consumption that way.  I didn't measure whether lowering clock rates gave significant power savings, but did observe that it does not hurt performance much.

In short, none of the settings produce nice smooth curves, it's all quite random, and this makes it really hard to find ideal settings.  From forum threads I get the feeling this is not normal for a 7970, most people (including ckolivas in the scrypt readme) seem to get more predictable performance changes and achieve better final results.

Also, have you tried Bitcoin mining with it - what performance do you get over there?
« Last Edit: April 12, 2013, 04:01:26 AM by gfoot »
BTC: 14ciemcbiGQ328NUTUz7j4wZ2tKy998C4a
LTC: LMUo6xEJZtPsNDpZe2QLEzmK1eW4tYZNLF

Offline fuzznarf

  • Jr. Member
  • **
  • Posts: 52
Re: xfx cards
« Reply #2 on: April 11, 2013, 03:21:29 PM »
thanks for the reply

on my gigabyte I am using
cgminer -o url:port -u uname -p pword --worksize 256 --thread-concurrency 8704 -g 2 --intensity 13
this gives me about 700ksh.

exact same settings, same computer, XFX card doing like 400-450khs.
i tried upping intensity to 14 and get HW errors, 12 gives about 300Khs
tried using shaders 2048 instead of thread-concurrency = 350ksh
tried lowering threads to 8192 = 350-400khs
i tried running both cards at the same time, still crappy performance on xfx
tried taking out my gigabyte card, still crappy performance.

this is driving me nuts...
LTC: LfvVoS9zGeRqXDyKpsvSQNXoyoNBgEeQc6
BTC: 1Bj1xQjQg8eJqZNkuozNEwLN58FYKqzxsz

Offline gfoot

  • Jr. Member
  • **
  • Posts: 165
Re: xfx cards
« Reply #3 on: April 11, 2013, 03:34:13 PM »
Please do try the settings I said in bold, especially "-g 9" coupled with low enough thread concurrency, and "powertune" set to 20%, as these are the main ones that improved it for me, otherwise I was stuck below 500kH/s.
BTC: 14ciemcbiGQ328NUTUz7j4wZ2tKy998C4a
LTC: LMUo6xEJZtPsNDpZe2QLEzmK1eW4tYZNLF

Offline gfoot

  • Jr. Member
  • **
  • Posts: 165
Re: xfx cards
« Reply #4 on: April 11, 2013, 07:29:30 PM »
Sorry, I've just realised that although these settings make cgminer claim to be hashing faster, it doesn't actually submit any shares.  I thought cgminer was different to reaper in that it checked that the stuff the GPU was feeding back made sense, but I guess not.  It is frustrating though that the statistics displayed are not measuring the rate of "correct" hashes calculated.  So I am still confused about which of the many KH/s-like numbers one should pay attention to.  It looks like maybe the only reliable test is to connect to a pool and watch the WU field.
BTC: 14ciemcbiGQ328NUTUz7j4wZ2tKy998C4a
LTC: LMUo6xEJZtPsNDpZe2QLEzmK1eW4tYZNLF

Offline fuzznarf

  • Jr. Member
  • **
  • Posts: 52
Re: xfx cards
« Reply #5 on: April 12, 2013, 12:21:50 PM »
so far the best I can do is 550-ish

ughh.

is anybody running an XFX 7970 over 650khs?
LTC: LfvVoS9zGeRqXDyKpsvSQNXoyoNBgEeQc6
BTC: 1Bj1xQjQg8eJqZNkuozNEwLN58FYKqzxsz

Offline Linus

  • Jr. Member
  • **
  • Posts: 52
Re: xfx cards
« Reply #6 on: April 12, 2013, 01:25:48 PM »
I havean XFX7970 GHz ed and so far the best I've gotten is 550 with a stable overclock.
Upwards of 600 overclocked, but causing the system to crash quite quickly.

Tried what people have posted here and in other threads, but no settings seem to give me more than 530-550.

I would appreciate if we could find out if it is at all possible to get them to those more interesting numbers :P
People seem to be posting their addresses in their signatures. I suppose I could do that too.
LTCiY4zuAj9yhediB7BDqSqy1VBfakz9ZP

Offline bruiser

  • Jr. Member
  • **
  • Posts: 180
Re: xfx cards
« Reply #7 on: April 12, 2013, 01:44:51 PM »
I haven't tried the XFX 7970's yet (and I probably wouldn't ever try any of XFX's stuff) but it might be a similar situation as the 7950's, where my Sapphires are significantly lower hash than MSI/Gigabyte. Almost 100kh/s per card, and I've probably spent 2 days worth of tinkering trying to get more and they just won't budge.

I'm not sure what you've tried, but here is what I'm running on my VisionTek 7970:

-I 13 -g 2 --threadconcurrency 8192 -w 256 --gpu-engine 1075  --gpu-memclock 1800   and that keeps me over 700kh/s constant while I'm using the desktop. I should note, this card is water cooled with a Hailea chiller--so it's probably not the norm.

I saw the same stuff as gfoot--basically, the overclocking didn't seem to help unless you got up in the 1700/1750+ range on the memory, and core overclocking does almost nothing for scrypt besides more heat--though a little is necessary to try to maintain the 0.57~ ratio.

Offline gfoot

  • Jr. Member
  • **
  • Posts: 165
Re: xfx cards
« Reply #8 on: April 12, 2013, 04:38:44 PM »
So on the power front, I found that the core voltage is semi-locked according to engine clock.  If you drop the engine clock to e.g. 800, you can drop the core voltage to maybe 1.00.  In fact setting powertune to negative values will change both those and the memory clock... it's weird how interlinked these settings are.

But who would buy such an expensive card if they cared about power use.

Anyway, I'm currently getting about 530KH/s reported, 470/m WU, on a P2Pool which is offering me difficulties around 200 on the cgminer scale (= 200/65536 in the Litecoin world).  It's not great but I can live with it.

I had some pollution of the AMD APP SDK, since I originally installed the latest version, then trusted the lower version's installer to overwrite it properly.  Since I cleaned it out of /opt and reinstalled it, my hash rates are unaffected but there are symptoms that show that at least something is different - I now have to set that GPU memory allocation environment variable, which I didn't seem to need before.
BTC: 14ciemcbiGQ328NUTUz7j4wZ2tKy998C4a
LTC: LMUo6xEJZtPsNDpZe2QLEzmK1eW4tYZNLF

Offline Floor

  • Newbie
  • *
  • Posts: 1
Re: xfx cards
« Reply #9 on: April 13, 2013, 05:57:25 PM »
Well im in the same boat and have 4 XFX dual fan 7970 cards, can only get them running stable around 530KH/s :(
Got a 5970 running at 690 and a gigabyte 7950 running at 620 in another rig without any problems but the XFX cards are a real pain to get them to atleast 600 :(

here is my config, my fourth card is crashing when I try a higher gpu / clock rate It might be a power issue so I ordered a powered riser and will see if that fixes the problem.

{

"intensity" : "13,13,13,13",
"gpu-engine" : "1051,1051,1051,960",
"gpu-fan" : "0-85,0-85,0-85,0-85",
"gpu-memclock" : "1744,1744,1744,1350",
"gpu-powertune" : "0,0,0,0",
"temp-cutoff" : "95,95,95,95",
"temp-overheat" : "85,85,85,85",
"temp-target" : "75,75,75,75",
"thread-concurrency" :"24576,24576,24576,16384",
"worksize" :"64,64,64,64",


"auto-fan" : true,
"expiry" : "120",
"failover-only" : true,
"gpu-threads" : "1",


"kernel-path" : "/usr/local/bin"
}