Glad I helped you out on the Litecoin Resource thread!
 
I see a problem in your screenshot though, its showing 89 - 90% which means you are getting lots iinvalid work turned in, which is not good. It should never be any lower then 99% (all 4 of my nvidia cards are staying above 99% at all times)
 
You have a stability problem somewhere still on the overclocks with your cards or the pool server you are using isn't very efficent, one of them is causing the % to be lower.
 

You should be setting the Titans to T14x32 and not T14x24.  Titans, 780, 780 Ti all can do 32 warps. You are limiting your self by only ruining 24 warps.

eduncan911
Oh, is that what those percentages mean?!?

You are right then. I'll mess with it more. Dang, though it was kicking butt.

either that or the ming pool server you were using causes the lower percentage as well (just found out about this today when I was testing other mining pool sites)

Spongebob28
You should be setting the Titans to T14x32 and not T14x24.  Titans, 780, 780 Ti all can do 32 warps. You are limiting your self by only ruining 24 warps.

I thought I read that was well!  But when running Autotune, I didn't see x32 as an option in the debug output.
 
Thanks!  I'm using T14x32 now, and...  Whoa...  VRAM memory usage has jumped to 4.3 GB now!
 

 
I am not really seeing any more hash though.  Maybe 5 to 10 at most... 
 

Spongebob28
Also, there is no point in using -C 2.  That command is ignored in the T kernel.  It says it at the top when you first start the miner if you have that command on. On a side note, you might want to use the -m command.   --->[-m] list of flags (0 or 1) to make the devices allocate their scrypt scratchpad in a single, consecutive memory block. On Windows Vista, 7/8 this may lead to a smaller memory size being used.

Could you explain this a bit more?  I thought the -C was implied with Titans (yes, ignored, didn't hurt to use it) and that when using -C that -m 0 was implied?  I'm still a bit confused to that as I didn't see any examples on how to use that...
 
When I specify:
 
cudaminer.exe -d 0,1 -i 0,0 -m 0,0 -l T14x32 -H 1
 
I get:
 
[2014-01-04 18:29:25] GPU #0: GeForce GTX TITAN with compute capability 3.5
[2014-01-04 18:29:25] GPU #0: the 'T' kernel requires single memory allocation
[2014-01-04 18:29:25] GPU #0: interactive: 0, tex-cache: 0 , single-alloc: 1
[2014-01-04 18:29:25] GPU #0: using launch configuration T14x32
[2014-01-04 18:29:25] GPU #1: GeForce GTX TITAN with compute capability 3.5
[2014-01-04 18:29:25] GPU #1: the 'T' kernel requires single memory allocation
[2014-01-04 18:29:25] GPU #1: interactive: 0, tex-cache: 0 , single-alloc: 1
[2014-01-04 18:29:25] GPU #1: using launch configuration T14x32

I don't know if that's right or not.

Set it to -m 1

-m 0 I believe turns it off.

For a 780, your setting should be T12x32

My 780 Classy liked T12x22 best (didn't try anytthing > x24 though, as I didn't know about x32 at the time).  I was getting up to 690 khash/s on my 780 Classified, but the power draw was insane - 850W from the wall with a single GPU!  We're talking a poor khash/J rating there.  It took a hell of an overclock at 1350 Mhz to get that high, not to mention a lot of voltage (1.35V, maxed in the Classified controller).  Anything higher started to get the "Could not validate on CPU" stuff, which I thought was my CPU overclock.
 
Didn't really spend more than an hour or so as I wanted to get my Titans back in.
 
The 780 Classy will be going into my Linux router in a few days, until the Step Up is ready to the 780 Ti (whcih I'll turn right around and sell New In Box, unopened, to pick up a couple ATI cards).  I think I liked it right around 620 khash/s for temps, power draw, etc.  Anything it took to get higher was just running into diminishing returns.
 
 

_Nite_
its not much better for me eduncan911 lol
 
GTX 770 = 220w @ 330 KH/s
GTX 570 = 180w @ 220 KH/s
GTX 570 = 180w @ 200 KH/s
GTX 460 = 150w @ 125 KH/s
 
so thats 730w for 875 KH/s
 
when I get the 3 x R9 270x it will basicly be double the hashrate with 1/3 less power usage. (probably like 1600 KH/s for 500w power consumption)

Are you undervolting?  That made a huge difference on my GTX 460 and Titans.  I figured about a 110W power draw from my EVGA GTX 460 EE at stock speeds but voltage turned all the way down.
 
Funny how the 780 Classy though didn't like any undervolting except at stock speeds.  Anything higher than stock wanted normal or higher voltage.  Maybe a sign that when you have that many phases (14 = 12+2), instead of like the Titan's 6 + 1 phases for the VRM section, you need normal voltage to drive them - and they will provide a lot of power.

_Nite_
Nah just running everything at out of the box settings

Except for the 780 Classy, which has a huge ACX cooler on it, everything I got (GTX 460 EE, 2x Titans) all overheated on "out of the box" settings.  Overheating = throttling or > 90C in my book.  Some fan settings helped, but not enough IMO.  
 
If you wanted to tweak only one thing, undervolting is what you would want to do.  The only GPU I ran into issues with was the GTX 460 undervolting it.  I had to bump it up 1 notch to be stable as it went too low.

Spongebob28
For a 780, your setting should be T12x32

So should I see better Hashrates when warps are optimized? I tried T12x32... and T15x32 nothing really changed.

_Nite_
yea well the 780 and Titan are hotter running cards so I casn understand that in your case
 
my GTX 770 and 2 x GTX 570 only hit 80C, and my GTX 460 hits 75C
 
None of them are throttling.

 
Actually, I found the sweet spot to get the Titans running at 79 to 80C: the Temp Limiter.  The 600/700 series have the famous Power Target that limits yep.  But there's also a Temp Limiter that you can set manually to control temps.  I was previously using the Power Target to limit fan/temps somewhat, but with the room temps drastically changing I was chasing a white ghost in that adventure.
 
For the last 9 hours, I've been running with fans at 100% and Target Temp set to 80C.  It's a night and day difference!
 
Current hash rates are:
 
[2014-01-05 09:34:52] GPU #0: GeForce GTX TITAN, 4071424 hashes, 659.27 khash/s
[2014-01-05 09:35:19] GPU #1: GeForce GTX TITAN, 18178048 hashes, 661.13 khash/s
[2014-01-05 09:35:19] accepted: 1045/1053 (99.24%), 1320 khash/s (yay!!!)
 
^- no more screenshots... I am logging to file now.  Btw, that was a bit difficult to figure out.  Previously I redirected the pipe with:
 
cudaminer -options >> logfile.txt
 
That didn't redirect the output of the cudaminer though.  The trick is to use "2>".  Like:
 
cudaminer -options 2>> logfile.txt
 
I've been coding for 25 years, and never saw that "2>" before.  Odd.
 
^- the hash rates fluctuate between 1250 and 1400, depending on the room temps.  If I crack a window (12F outside) it shoots to 1405 khash/s @ 1202 Mhz core, and 100% fan actually cools things down to 62C!  But, the boss didn't like that (wife). 
 
^- this also allows me to do exactly what I want to do during the day: limit noise!  I can lower the fan speed manually to an acceptable 75%.  The temps stay at 80C, but my hash rates drops to about 1000 khash/s.  I got the sound where I want it.  Temps where I want it.  Easily by setting the Temp Target and forgetting it.  And, my hash rates increased from 800 to 1000 khash/s with this combination.  Before, the wild core clock fluctuations with the Power Target limiting I think was causing more harm - and I couldn't control the temps long term.
 
I contribute the "drops", or boos, to the wild fluctuations with the GPU usage and voltages up and down (goes even lower now with Temp Limiting).  But still up at 99% since I started it at 80C temp limit.
 

 
Using the temp limiter made the core speeds failr high and stable now.  Only the voltages fluctuate a little (to lower temps), and the GPU usage seems to be more erratic.  
 
And this is in a non-vented room.  As of a few minutes ago, I am using a fan now to blow the heat out into the hallway in the upstairs from my office to see if I can lower the temps even more.
 

shadowboricua

eduncan911
My 780 Classy liked T12x22 best (didn't try anytthing > x24 though, as I didn't know about x32 at the time).  I was getting up to 690 khash/s on my 780 Classified, but the power draw was insane - 850W from the wall with a single GPU!  We're talking a poor khash/J rating there.  It took a hell of an overclock at 1350 Mhz to get that high, not to mention a lot of voltage (1.35V, maxed in the Classified controller).  Anything higher started to get the "Could not validate on CPU" stuff, which I thought was my CPU overclock.
 
Didn't really spend more than an hour or so as I wanted to get my Titans back in.
 
The 780 Classy will be going into my Linux router in a few days, until the Step Up is ready to the 780 Ti (whcih I'll turn right around and sell New In Box, unopened, to pick up a couple ATI cards).  I think I liked it right around 620 khash/s for temps, power draw, etc.  Anything it took to get higher was just running into diminishing returns.

Just tried T12x22 on my GTX 780 SC ACX; hashrates went up to 500! I keep wondering if there are even better configurations. But how to find them?

 
Use the options: -l auto -D --benchmark
 
-l (as in "L", or Launch Config) is how you are specifying the T12x22.  -l T12x22, per device.  If you have multiple devices, you can fine tune each one separated by a comma: -l T12x22,T15x32
 
-D, not -d, turns on Debugging - with exactly what you are looking for (T1-T256 or something, and x1=x24).  It doesn't show up to x32 though, hence why I wasn't using that at first.  
 
The readme file is correct in that you should only use -D for a single GPU.  If you have multiple devices, only configure it
to look for 1 device at a time to fine tune exactly the best config (and it will flood you with multiple devices if you don't).
 
So to recap, use the following if you have multiple GPUs (aka devices):
 
GPU #0: cudaminer -d 0 -l auto -D --benchmark
GPU #1: cudaminer -d 1 -l auto -D --benchmark
GPU #2: cudaminer -d 2 -l auto -D --benchmark
 
If you only have a single GPU, you can omit the -d # part.
 
And I set my console window to be 200 width in order to see the entire table.  Go to the upper-left corner, click it, and then Properties.  Set the width to 200 as well as the size to 200.  Then run it again.
 
In my opinion, the x## number has a direct correlation to the amount of ram it uses.  For me, x22 filled up 2.8GB on my Titan and 780 Classy.  x24 hit 3.00 GB on my Titan and Classy, x32 hits 4.3 GB on my Titan - and I haven't tried it on my 780 Classy, etc.  In other words, I think x24 would fill up the 3GB on your 780 which may be too much memory pressure for the best benchmark.  
 
x22 seemed to be the sweet spot with 384-bit bus memory limited to 3GB of ram.  I am using x32 on my Titans, and not really seeing any difference in hashes.  Well, maybe a little.
 
You have to remember that the 384-bit bus is limited by bandwidth - you won't ever fill up the entire ram of 6GB on a Titan with a 384-bit bus, unless you have a slow computation (gaming and mining is not slow).   But, you can fill up 3GB on a 384-bit bus just fine. On a 256-bit bus, like the 770, 600 series, etc, you can't ever use the entire 4 GB on say my 670s I had, or a 770, etc.  Same limitation of bandwidth.  But you can fill up 2 GB on a 256-bit bus just fine.  
 
The whole "double memory" thing helps, yes (256-bit double memory to 4 GB, and 384-bit double memory pf 6 GB only in the TItan).  But, not for the entire memory amount.  I would safely say expect to use only 70 to 75% of the memory of when it is doubled, limited by bandwidth.  This may be why my Titans are only using 4.3 GB, instead of 4.5 GB like it should scale between x22, x24 and x32.

With the setting of T12x13, my 780 Classified only uses 1.8 GB.  My 670 SCs are using 1.1 GB and 995 MB with the setting of K7x32.

Good question, Nite. 

wow that benchmark feuture is neat I didn't even know it had that, my new launch value is K24x16 now and using 1.9GB of VRAM
 
performance didn't change much though and now my desktop is super laggy..... lol
 
I'm gonna try this on my GTX 570 cards since I don't use those computers, but on my GTX 770 I will have to enable interactive mode.

Use --benchmark with -D (capital D) and "auto". That's the debug that will go through all iterations for you. I ran it 2 or three times to find the most repetitive ones. Then I fine tune along the horizontal line it liked the best with the x##.

Make sure to set your console to 200 width (upper-left corner click, properties and set width and size to 200).