Topic: GPUs can't run at 20 intensity..?? (Read 2849 times)

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, all of which 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.

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.


DRIVERS AND OPENCL SDK

The choice of driver version for your GPU is critical, as some are known to
break scrypt mining entirely while others give poor hashrates. As for the
OpenCL SDK installed, for AMD it must be version 2.6 or later.


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.

On windows the same commands can be passed via a batch file if the following
lines are in the .bat before starting cgminer:
setx GPU_MAX_ALLOC_PERCENT 100
setx GPU_USE_SYNC_OBJECTS 1

--intensity XX (-I 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. High intensities
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. Note that if you do NOT specify an
intensity, cgminer uses dynamic mode which is designed to minimise the harm
to a running desktop and performance WILL be poor. The lower limit to intensity
with scrypt is usually 8 and cgminer will prevent it going too low.
SUMMARY: Setting this for reasonable hashrates is mandatory.

--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. If this is not set, cgminer will query the
device for how much memory it supports and will try to set a value based on
that.
SUMMARY: This will get you started but fine tuning for optimal performance is
required.

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.


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

--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 and is ultimately a BETTER way to tune performance.
SUMMARY: Spend lots of time finding the highest value that your device likes
and increases hashrate.

-g:
Once you have found the optimal shaders and intensity, you can start increasing
the -g value till cgminer fails to start. This is really only of value if you
want to run low intensities as you will be unable to run more than 1.
SUMMARY: Don't touch this.

--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.
SUMMARY: Don't touch this.


Related parameters:
--worksize XX (-w XX)
Has a minor effect, should be a multiple of 64 up to 256 maximum.
SUMMARY: Worth playing with once everything else has been tried but will
probably do nothing.

--vectors XX (-v XX)
Vectors are NOT used by the scrypt mining kernel.
SUMMARY: Does nothing.


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.

Third, the combination of motherboard, CPU and system ram ALSO makes a
difference, so values that work for a GPU on one system may not work for the
same GPU on a different system. A decent amount of system ram is actually
required for scrypt mining, and 4GB is suggested.

Finally, the power consumption while mining at high engine clocks, very high
memory clocks can be far in excess of what you might imagine.
For example, a 7970 running with the following settings:
--thread-concurrency 22392 --gpu-engine 1135 --gpu-memclock 1890
was using 305W!

---
TUNING AN AMD RADEON 7970
Example tuning a 7970 for Scrypt mining:

On linux run this command:
export GPU_MAX_ALLOC_PERCENT=100
or on windows this:
setx GPU_MAX_ALLOC_PERCENT 100
in the same console/bash/dos prompt/bat file/whatever you want to call it,
before running cgminer.

First, find the highest thread concurrency that you can start it at. They should
all start at 8192 but some will go up to 3 times that. Don't go too high on the
intensity while testing and don't change gpu threads. If you cannot go above
8192, don't fret as you can still get a high hashrate.

Delete any .bin files so you're starting from scratch and see what bins get
generated.

First try without any thread concurrency or even shaders, as cgminer will try to
find an optimal value
cgminer -I 13

If that starts mining, see what bin was generated, it is likely the largest
meaningful TC you can set.
Starting it on mine I get:
scrypt130302Tahitiglg2tc22392w64l8.bin

See tc22392 that's telling you what thread concurrency it was. It should start
without TC parameters, but you never know. So if it doesn't, start with
--thread-concurrency 8192 and add 2048 to it at a time till you find the highest
value it will start successfully at.

Then start overclocking the eyeballs off your memory, as 7970s are exquisitely
sensitive to memory speed and amazingly overclockable but please make sure it
keeps adequately cooled with --auto-fan! Do it while it's running from the GPU
menu. Go up by 25 at a time every 30 seconds or so until your GPU crashes. Then
reboot and start it 25 lower as a rough start. Mine runs stable at 1900 memory
without overvolting. Overvolting is the only thing that can actually damage your
GPU so I wouldn't recommend it at all.

Then once you find the maximum memory clock speed, you need to find the sweet
spot engine clock speed that matches it. It's a fine line where one more MHz
will make the hashrate drop by 20%. It's somewhere in the .57 - 0.6 ratio range.
Start your engine clock speed at half your memory clock speed and then increase
it by 5 at a time. The hashrate should climb a little each rise in engine speed
and then suddenly drop above a certain value. Decrease it by 1 then until you
find it climbs dramatically. If your engine clock speed cannot get that high
without crashing the GPU, you will have to use a lower memclock.

Then, and only then, bother trying to increase intensity further.

My final settings were:
--gpu-engine 1157  --gpu-memclock 1900 -I 20
for a hashrate of 725kH.

Note I did not bother setting a thread concurrency. Once you have the magic
endpoint, look at what tc was chosen by the bin file generated and then hard
code that in next time (eg --thread-concurrency 22392) as slight changes in
thread concurrency will happen every time if you don't specify one, and the tc
to clock ratios are critical!

Good luck, and if this doesn't work for you, well same old magic discussion
applies, I cannot debug every hardware combo out there.

Your numbers will be your numbers depending on your hardware combination and OS,
so don't expect to get exactly the same results!