Scrypt mining, AKA litecoin mining, for GPU is completely different to sha256
used for bitcoin mining. The algorithm was originally developed in a manner
that it was anticipated would make it suitable for mining on CPU but NOT GPU.
Thanks to some innovative work by Artforz and mtrlt, this was proven to be
wrong. However, it has very different requirements to bitcoin mining and is a
lot more complicated to get working well. Note that it is a ram dependent
workload, and requires you to have enough system ram as well as fast enough
GPU ram. If you have less system ram than your GPU has, it may not be possible
to mine at any reasonable rate.

There are 5 main parameters to tuning scrypt, 2 of which you MUST set, and
the others are optional for further fine tuning. When you start scrypt mining
with the --scrypt option, cgminer will fail IN RANDOM WAYS. They are all due
to parameters being outside what the GPU can cope with. Not giving cgminer a
hint as to your GPU type, it will hardly ever perform well.

NOTE that if it does not fail at startup, the presence of hardware errors (HW)
are a sure sign that you have set the parameters too high.


Step 1 on linux:
export GPU_MAX_ALLOC_PERCENT=100
If you do not do this, you may find it impossible to scrypt mine. You may find
a value of 40 is enough and increasing this further has little effect.

export GPU_USE_SYNC_OBJECTS=1
may help CPU usage a little as well.

--shaders XXX

is a new option where you tell cgminer how many shaders your GPU has. This
helps cgminer try to choose some meaningful baseline parameters. Use this table
below to determine how many shaders your GPU has, and note that there are some
variants of these cards, and nvidia shaders are much much lower and virtually
pointless trying to mine on.

GPU Shaders
7750 512
7770 640
7850 1024
7870 1280
7950 1792
7970 2048

6850 960
6870 1120
6950 1408
6970 1536
6990 (6970x2)

6570 480
6670 480
6790 800

6450 160

5670 400
5750 720
5770 800
5830 1120
5850 1440
5870 1600
5970 (5870x2)

These are only used as a rough guide for cgminer, and it is rare that this is
all you will need to set.


--intensity XX

Just like in bitcoin mining, scrypt mining takes an intensity, however the
scale goes from 0 to 20 to mimic the "Aggression" used in mtrlt's reaper. The
reason this is crucial is that too high an intensity can actually be
disastrous with scrypt because it CAN run out of ram. Intensities over 13
start writing over the same ram and it is highly dependent on the GPU, but they
can start actually DECREASING your hashrate, or even worse, start producing
garbage with HW errors skyrocketing. The low level detail is that intensity is
only guaranteed up to the power of 2 that most closely matches the thread
concurrency. i.e. a thread concurrency of 6144 has 8192 as the nearest power
of two above it, thus as 2^13=8192, that is an intensity of 13.


Optional parameters to tune:
-g, --thread-concurrency, --lookup-gap

-g:
Once you have found the optimal shaders and intensity, you can start increasing
the -g value till cgminer fails to start. Rarely will you be able to go over
about -g 4 and each increase in -g only increases hashrate slightly.

--thread-concurrency:
This tunes the optimal size of work that scrypt can do. It is internally tuned
by cgminer to be the highest reasonable multiple of shaders that it can
allocate on your GPU. Ideally it should be a multiple of your shader count.
vliw5 architecture (R5XXX) would be best at 5x shaders, while VLIW4 (R6xxx and
R7xxx) are best at 4x. Setting thread concurrency overrides anything you put
into --shaders.

--lookup-gap
This tunes a compromise between ram usage and performance. Performance peaks
at a gap of 2, but increasing the gap can save you some GPU ram, but almost
always at the cost of significant loss of hashrate. Setting lookup gap
overrides the default of 2, but cgminer will use the --shaders value to choose
a thread-concurrency if you haven't chosen one.


Overclocking for scrypt mining:
First of all, do not underclock your memory initially. Scrypt mining requires
memory speed and on most, but not all, GPUs, lowering memory speed lowers
mining performance.

Second, absolute engine clock speeds do NOT correlate with hashrate. The ratio
of engine clock speed to memory matters, so if you set your memory to the
default value, and then start overclocking as you are running it, you should
find a sweet spot where the hashrate peaks and then it might actually drop if
you increase the engine clock speed further. Unless you wish to run with a
dynamic intensity, do not go over 13 without testing it while it's running to
see that it increases hashrate AND utility WITHOUT increasing your HW errors.