7750 mining...maybe not so meh after all

[Isepick]

Ok important stuff 1st:

max OC @stock volts so far w/ cgminer: -I 5 -w 256
950/1010/1.1 volts 110 MHash/s at 90 watts (90-55)=35 watts for 110/35= 3.14 Mhash/w    (U=1.62/m)

voltage appears to be locked in the bios to certain speeds. I could not get the voltage to drop until I underclocked the core clock to 400 Mhz.
At 400/1010/.95 volts gave me 50 Mhash using 15 watts for 50/15= 3.33 Mhash/w

I don't think the #'s are going to change appreciably until the gcn kernels get better or the tools to edit the new 7xxx bios appear.


Card is an Asus HD-7750-0-1-GD5 $109 at Newegg

system idle: 55 watts

using the catalyst 12.3 drivers (aka 8.95RC) with phoenix 1.7.5 with diakgcn kernel. aggression=12 vectors2 and worksize=256 (128 lost a few Mhash).
system mining:
stock 820/1150/1.1 volts   103 Mhash/s at 125 watts  (125-55)= 70 watts mining for 103/70= 1.476 Mhash/w

max oc with garbage Asus GPUTweak utility:
870/1010/1.13 volts    109 Mhash/s at 130 watts (130-55) = 75 watts for 109/75= 1.45 Mhash/w

max underclock with crappy Asus util:
770/1010/1.09 volts     97 Mhash/s at 120 watts (120-55)= 65 watts for 97/65= 1.49 Mhash/w

system is a clean install of win7 32 bit. MSI Afterburner is useless for anything except graphs. Trixx is even more useless for this card. Asus utility has a very narrow OC/UC range and the sliders for voltage/gpu clock are annoyingly locked together.

Fan speed never goes over 40% even at max clocks, temp never goes over 65C. If set manually, the fan isnt audible until >60%. System is caseless.

EDIT:

I got cgminer working!! I had to remove the app sdk that 12.3 installed and then go delete amdocl.dll and opencl.dll in win/system32 and then install the 2.6rc3 opencl. This gave me the magic 831 that let cgminer build its bins. Latest diakgcn kernel (diakgcn120216.cl) didn't find any shares tho, it just sat there showing Mhashes, but not actually finding anything. So I am using the default poclbm kernel.

Much better Mhash/w numbers and superior overclocking control.

Gonna let cgminer run at those clocks at the top and go get some sleep. Will tweak some under/overvolting tomorrow. I have no idea why power usage is that drastically different for the 2 programs. I even rebooted the machine and restarted cgminer just to be sure everything was right.

[1715408574]

1715408574

[1715408574]

1715408574

[1715408574]

1715408574

[luo demin]

Do you think more optimized drivers will help these cards or is teh lack of SP's whats giving these low numbers ?

[Isepick]

Only 512 SPs, so I wasn't expecting a huge amount of Mhashes. I mainly got this card to see where it would fall on the Mhash/w scale. Better drivers and mining kernels will improve the numbers further. At .0802 bitcoins per day, it is going to take a while to break even tho 

[luo demin]

At least you can recoup most your cost if you sell it thats the best thing about new hardware !

[jake262144]

Nice findings, thanks for sharing OP.

Properly tuned, GCN cards can get dangerously close to BFL unicorn FPGA with the benefits of having a full manufacturer's warranty and support and a non-zero resale value.

Personally, I'm waiting for the 7990 cards - they will be using top-quality bin chips so the power efficiency should rock.
The initial price will be ungainly but a month of competition from nVidia's Kepler should drop it to an acceptable level.

[Mousepotato]

Ok important stuff 1st:

max OC @stock volts so far w/ cgminer: -I 5 -w 256
950/1010/1.1 volts 110 MHash/s at 90 watts (90-55)=35 watts for 110/35= 3.14 Mhash/w    (U=1.62/m)

That's still 2.5X MH/w less than a 5970 using the same metrics (5.67 MH/w).  But for gross power consumption, that 7750 appears very fuel efficient, so to speak.

[racerguy]

I think it would compete handsomely in mhash/w with an undervolt to 0.9v.  Probably superior to the 5970.

[Isepick]

Voltage appears tied to clock speed in the bios. I couldn't drop the voltage until I dropped the core clock to 400 Mhz. Until we get some bios editing tools or the coders improve their kernels we probably won't get better Mhash/w numbers.