3x7970 Mining Results.

[DeathAndTaxes]

Is there any table or something where we can see the performance for a 5870, 5970 etc. with Vanitygen ?

Maybe we can add these results to that table / page ?

My 5870 gets around 31 Mkeys/s with 935 core clock: https://bitcointalk.org/index.php?topic=25804.msg610714#msg610714
Don't underclock the memory when using vanitygen, it reduces the key rate drastically.

wow good tip.  Didn't think of that.

On edit: looks like vanitygen is heavily mem clock limited.

You get 31 MK/s @ 935 core and I get 20 MK/s @ 775 core (5870 and 5970 have same core).  Your core is 30% higher but you get >50% higher throughput which makes me think that even @ 1000 MHz memclock is the bottleneck. 

Just wonder what memclock are you running at?

Also sorry 7970 to push this off topic but maybe it helps in getting best throughput on a 7970 also.

[1713438655]

1713438655

[1713438655]

1713438655

[DiabloD3]

Someone with a 7970, please measure the exact length of the card. My Gigabyte Super-5850 is on a 5870 carrier and just BARELY fits, it rubs against the ass end of the harddrives. I wanna make sure a 7970 will actually fit.

[runeks]

Just wonder what memclock are you running at?

I'm going by memory (no pun intended), but I'm pretty sure I upped the memory clock to its max when I found out it was heavily memory dependent, so that would be 1300 MHz for my 5870.

[sveetsnelda]

Someone with a 7970, please measure the exact length of the card. My Gigabyte Super-5850 is on a 5870 carrier and just BARELY fits, it rubs against the ass end of the harddrives. I wanna make sure a 7970 will actually fit.

I'll gladly measure mine when I get back from work, but they all seem to slightly vary by brand/model (even though they are reference designs).  Anandtech has a good chart though:

http://www.anandtech.com/show/5319/amd-radeon-hd-7970-launch-recap

[celcoid]

Someone with a 7970, please measure the exact length of the card. My Gigabyte Super-5850 is on a 5870 carrier and just BARELY fits, it rubs against the ass end of the harddrives. I wanna make sure a 7970 will actually fit.

11.2in

[ArtForz]

linux, cat 11.12, sdk2.6, vanitygen git, disabled BFI_INT to get correct results
925/1375 28.1Mk/s
1125/1375 32.4Mk/s
1125/1575 33.2Mk/s
1170/1600 34.1Mk/s

some more testing, disabled EXPENSIVE_BRANCHES and DEEP_VLIW, -g 2048x2048 -b 256
925/1375 37.1Mk/s
1125/1375 43.0Mk/s
1125/1575 43.7Mk/s
1170/1600 45.0Mk/s

more tweaking...
925/1375 39.8Mk/s
1125/1375 46.6Mk/s
1170/1600 49.1Mk/s

edit: 50.1Mk/s at 1200/1600

[DeathAndTaxes]

1170/1600 49.1Mk/s

Almost 50MK/s from a single GPU.  I got so say I am impressed.  The memory throughput on 7970 really shines.

BTW: Those tweaks and adjustments are they possible on 5000 series cards or was it some 7970 model specific optimizations?

[ArtForz]

Doubt it, basically all I did was disable the ati-specific tweaks, replaced the BFI_INT parts with bitselect() and set fixed work group sizes (16x16 for first and 3rd kernel, 64 for 2nd)

[deepceleron]

oclvanitygen -d 0 1Bitcoinz    (0 = device # and the string is the address searching for 1Bitcoinz is impossibly long which prevent the generator from getting lucky too soon)

<--- not impossibly long...

Another thing if you are running oclvanitygen, it will benefit from memory as overclocked as you can go.

[ArtForz]

Well, at least on 7970 it seems not horribly memory limited.
925 to 1125 = +21.6% core gives +17.1% performance (= 80% effective)
even at 1125 core 1375 to 1575 memclock =+14.5% only gets ~ +2.5% performance (= ~18% effective)
That looks a lot more core than memory limited to me.
So if it's *supposed* to be mem limited... that kernel still has lots of room for improvements on 79xx.

[DeathAndTaxes]

Well, at least on 7970 it seems not horribly memory limited.
925 to 1125 = +21.6% core gives +17.1% performance (= 80% effective)
even at 1125 core 1375 to 1575 memclock =+14.5% only gets ~ +2.5% performance (= ~18% effective)
That looks a lot more core than memory limited to me.
So if it's *supposed* to be mem limited... that kernel still has lots of room for improvements on 79xx.

Well 79xx series uses 384 bit memory bus so clock for clock it has 50% more memory throughput than the 6970.  Nothing like a brute force 50% boost to solve a bottleneck.

[jjiimm_64]

so what was the top hash on this card, regardless of efficiency?  if power is not problem?

someone needs to add to  https://en.bitcoin.it/wiki/Mining_hardware_comparison

[Syke]

And for those of you with cheap electricity:

2486 Mhash/sec @ 793 watts -- 3.13Mhash/watt  (1056mV, 1075Mhz/200Mhz).  Stable.

That's with 4 cards, so 621 MH/s per card.

[jjiimm_64]

well, we will see, I just bot 4 cards yesterday

[deepceleron]

well, we will see, I just bot 4 cards yesterday

I would have purchased them manually....

$2000 of the most expensive mining cards at a high difficulty = a loooong position.

[jjiimm_64]

well, we will see, I just bot 4 cards yesterday

I would have purchased them manually....

$2000 of the most expensive mining cards at a high difficulty = a loooong position.

aggreed. 

Difficulty Factor   1250757.73927
Hash Rate (mega-hashes / second)     2500
Exchange Rate ($/฿)   6.8

   Coins   Dollars
per Day   ฿2.01   $13.67
per Week   ฿14.07   $95.70
per Month   ฿61.12   $415.59


5 month payoff.  ......maybe.

[Syke]

per Month   ฿61.12   $415.59

5 month payoff.  ......maybe.

Plus, great resale value, makes it a pretty good deal.

[yochdog]

Here are some results so far on a dedicated miner:

ASRock 870 Extreme 3 R2.0
Athlon II X3 455
Two cores disabled
CPU voltage dropped by 0.1V
1 stick of DDR3 1066
Seasonic X-1250 PSU (80 Plus Gold)
Windows 7 64 bit
60GB Vertex 2 SSD

All of my dedicated miners are usually using a headless Ubuntu install on a flash drive, but since there probably aren't any programs for Linux that will allow me to change the clocks/voltage on these cards yet (besides what's allowed in aticonfig), I'm using Windows to test this.

2202 Mhash/sec @ 583 watts -- 3.77Mhash/watt  (962 mV, 975Mhz/160Mhz)
2106 Mhash/sec @ 553 watts -- 3.80Mhash/watt  (949 mV, 925Mhz/160Mhz)
2134 Mhash/sec @ 528 watts -- 4.04Mhash/watt  (931 mV, 925Mhz/160Mhz)  (one of my cards will not stay stable at this voltage though.  The other 3 are solid.)
2060 Mhash/sec @ 498 watts -- 4.13Mhash/watt  (918 mV, 900Mhz/160Mhz)  (all cards stable)

If you still think these cards aren't efficient, you're drunk.  If they were still using VLIW, they'd be insane at hashing.  All GPUs show 99 percent utilization.  If small improvements can be made to the miner kernels, efficiency will be further improved.

I've got a small external fan blowing on the rig (as with all of my mining rigs).  The temps on these cards while mining at 2060 Mhash/sec are 63/64/65/61C.  The fan speeds on auto are adjusting to 30/30/30/27% respectively.

All wattages were read with a Kill-a-watt (which makes the measurements on the AC side, of course).

Any idea how this compares on a mh/watt basis with 6970's?  I am in the process of replacing a lot of 6970's, and would love to get an idea of how much more efficient these are.

[plastic.elastic]

Here are some results so far on a dedicated miner:

ASRock 870 Extreme 3 R2.0
Athlon II X3 455
Two cores disabled
CPU voltage dropped by 0.1V
1 stick of DDR3 1066
Seasonic X-1250 PSU (80 Plus Gold)
Windows 7 64 bit
60GB Vertex 2 SSD

All of my dedicated miners are usually using a headless Ubuntu install on a flash drive, but since there probably aren't any programs for Linux that will allow me to change the clocks/voltage on these cards yet (besides what's allowed in aticonfig), I'm using Windows to test this.

2202 Mhash/sec @ 583 watts -- 3.77Mhash/watt  (962 mV, 975Mhz/160Mhz)
2106 Mhash/sec @ 553 watts -- 3.80Mhash/watt  (949 mV, 925Mhz/160Mhz)
2134 Mhash/sec @ 528 watts -- 4.04Mhash/watt  (931 mV, 925Mhz/160Mhz)  (one of my cards will not stay stable at this voltage though.  The other 3 are solid.)
2060 Mhash/sec @ 498 watts -- 4.13Mhash/watt  (918 mV, 900Mhz/160Mhz)  (all cards stable)

If you still think these cards aren't efficient, you're drunk.  If they were still using VLIW, they'd be insane at hashing.  All GPUs show 99 percent utilization.  If small improvements can be made to the miner kernels, efficiency will be further improved.

I've got a small external fan blowing on the rig (as with all of my mining rigs).  The temps on these cards while mining at 2060 Mhash/sec are 63/64/65/61C.  The fan speeds on auto are adjusting to 30/30/30/27% respectively.

All wattages were read with a Kill-a-watt (which makes the measurements on the AC side, of course).

Any idea how this compares on a mh/watt basis with 6970's?  I am in the process of replacing a lot of 6970's, and would love to get an idea of how much more efficient these are.

If those numbers are correct, these cards are double in term of efficiency.

[sveetsnelda]

I haven't tried to underclock any of my 6970's, so I am unsure.  Before I sold my XFX reference boards, I was getting 1580Mhash @ 683W with a slight overclock and a decent voltage decrease (900/340Mhz at 1.050V).  This was not on the same power supply as my 7970 tests (this PSU was slighly less efficient, but not much).  The rest of the components were the same though.

My non-reference 6970s (MSI Lightning) are using 25 watts more per card -- even at the same voltage.