Hello everyone,
because our beloved mods don't get that some people just dont want to talk much and only post when needed (yeah, it's a paradox isn't it - not to want to post junk on forums) I will post in here, when actually this should go to the cgminers out there (in the mining software sub-forum)
I run a Sapphire 6870 FleX 1GB DDR5 card and after having some hiccups I went onwards to fully optimize and balance my LTC(please mind i said LTC) generation.
My OS is Win 7 Ultimate 64 Bit with 4 GB of system RAM.
Because this is also my main PC, i cannot run cgminer all the time at the same settings, so I basically need 3:
- Sleep profile - low noise, close to 70C (I am cautious) but as high hash rate as possible
- Desktop profile - noise not important - close to 78C - high hash - but running in "I dont even notice something is running" mode
- Performance mode - noise absolutely not important (but i don't want my fan to die on me so no more than 85%) - temp not higher than 78C and...you guessed it...max hash rates! - when I'm away from home
Okay, now I know that max noise of the fan i can call silent is 53% so for the sleep profile thats max.
I have 1120 shaders and cgminer help says i should use 4x shaders for vliw5 so 4480....
And now my testing (done by letting the values stabilize on the same pool - so same difficulty):
thread concurrency 8192 - worksize 256
Intensity 17: 287.9 kH/s -- 82% fan speed -- 76.0C (target 78C)
Intensity 16: 282.0 kH/s -- 81% fan speed -- 76.0C (target 78C)
Intensity 15: 271.8 kH/s -- 81% fan speed -- 75.5C (target 78C)
thread concurrency 8192 - worksize 128
Intensity 17:
299.0 kH/s -- 85% fan speed -- 76.5C (target 78C) -FASTEST config that works so far
Intensity 16: 296.7 kH/s -- 84% fan speed -- 76.5C (target 78C)
Intensity 15: 286.7 kH/s -- 84% fan speed -- 76.0C (target 78C)
thread concurrency 4480 - worksize 256
Intensity 17: 288.7 kH/s -- 80% fan speed -- 77.0C (target 78C) - FASTER and cooler than 8192 w256 at same Intensity
Intensity 16: 282.7 kH/s -- 79% fan speed -- 76.5C (target 78C) - FASTER and cooler than 8192 w256 at same Intensity
Intensity 15: 272.6 kH/s -- 78% fan speed -- 76.0C (target 78C) - FASTER and cooler than 8192 w256 at same Intensity
thread concurrency 4480 - worksize 128
Intensity 17:
301.7 kH/s -- 81% fan speed -- 75.0C (target 78C) - HOLY CRAP - fast and cool but look below
Intensity 16:
298.7 kH/s -- 81% fan speed -- 75.0C (target 78C) - HOLY CRAP - fast and cool but look below
Intensity 15:
287.9 kH/s -- 77% fan speed -- 75.5C (target 78C) - HOLY CRAP - fast and cool but look below
you would expect it to be faster and cooler than the 8192 w128 if you look at the worksize-thread concurrency comparison, and IT IS...but
YOU GET HARDWARE ERRORS. Howcome? don't ask me....ckolivas or someone who understands it...please enlighten me...
Howcome at higher worksizes (256) everything works smoothly and at smaller,easier workloads (128) you get hardware errors i dont know.
What to do to get rid of them..considering that at tc 4480 everything is a wee bit faster and LOTS cooler.
UPDATE: - new settings tested with very good results - sadly temp wise is not comparable because of much hotter environment
- also decided to completely skip worksize 256 because of past results
thread concurrency 5632 - worksize 128
Intensity 17:
301.8 kH/s - Faster than any previous config at same intensity
Intensity 16:
298.8 kH/s - Faster than any previous config at same intensity
Intensity 15:
287.9 kH/s - Faster than any previous config at same intensity
UPDATE 2: After extensive reading of this thread:
https://bitcointalk.org/index.php?topic=92887And after Tittiez's discoveries:
https://bitcointalk.org/index.php?topic=92887.msg1045832#msg1045832I started to understand a little more the correlation between thread concurrency and worksize so here are my (partial) discoveries:
1. at the same worksize, decreasing thread concurrency increases speed - at some point you will reach the limit of your hardware (in my case that was tc 4480 w128)
2. at smaller worksizes you need a higher level of thread concurrency to keep yourself over the limit of the hardware (DO NOT FORGET 1.)
3. thread concurrency works best if it's a multiple of the shader count of your card (thus for me that means 1120, 2240, 3360, 4480, 5600, 6720, 7840 - 8960 will result in buffer error, need to test some more)
4. it seems that smaller worksizes increase hashing speed (the concept resembles to a funnel: if you try to send too much through a funnel it will overflow and will make it harder for the liquid to get through, yet if you get to the point where you just barely send enough to let it flow unhindered, you will get a far higher rate of fluid through that funnel)
Considering what I found out, here are my newest results
thread concurrency 6720 - worksize 64 - memory 900 mHz (default:1050 mHz)
(tc 5600 gives HW errors see points 2 and 3 and FOR THIS CARD lower memory speed means lower temps and about 1 kH/s extra)
Intensity 17:
302.1 kH/s -- 74% fan speed -- 77.5C (target 78C) - Even faster and cooler than ever before!
Intensity 16:
297.8 kH/s -- 74% fan speed -- 77.0C (target 78C)
Intensity 15:
292.8 kH/s -- 74% fan speed -- 77.0C (target 78C) - Even faster and cooler than ever before!
NOTE: stale shares were never my concern, with any previous config I never got more than 1.5%-2%, YET for some reason WITH THIS CONFIG ABOVE
I get a superb speed and temp AND only ~0.5% stales (so about 1.5% faster speed than these speeds, by submitting more valid shares)
The quest for the best optimized 6870 continues....
