CCminer(SP-MOD) Modded GPU kernels.

[myagui]

See if I can post it tomorrow...but Scrypt-Jane:15 I get 1.24 mh per 750ti if that's any help for now.

Unfortunately that doesn't help, as I really don't recall Scrypt-Jane:15 performance on the 750ti at all (and neither can I test it). All of my performance & test references are on Scrypt-Jane:16 specifically. Thanks for posting though!

0.38 khash (750ti 2GB, win 7)
but only 1.05 GB vram is in use ...

I'm particularly curious about Vista, as memory management differences might explain better performance on that one.

[...] So a 750ti with 2 GB is hashing as fast as a 750ti 4 GB at scrypt-n (chacha) algos.
(970 4 GB allocates a max of 2.1 GB, above that = performance drop) [...]

I find that statement to be true for Windows (7 & 8, still dunno about Vista), but not for Linux. On Linux, you can use different launch settings for the 4GB cards and much better performance can be achieved. My 980's run fastest with about 3.4 GB of VRAM allocated, but only on Linux. Might relate to the memory backoff settings discussed up-thread.

device_backoff = is_windows() ? 12 : 2;
Windows backoff 12%, non-Windows backoff 2%

device_backoff = is_windows() ? 1 : 2;
Windows backoff 1%, non-Windows backoff 2%

I suppose this is the change. Have not done this modification myself before though.

[sp_]

Edit: 200 pages! Congrats and keep up the good work sp_ (sent a few beers your way).

Thanks alot

[Epsylon3]

the interactive option also is made for that... i guess interactive should be checked instead of is_windows()

[sp_]

Here is a tip for you. Delete the error.log file in the ccminer folder, it can get huge after a while.

[Fuzzbawls]

the interactive option also is made for that... i guess interactive should be checked instead of is_windows()

Would love to see this change made! I was getting quite frustrated with mining scrypt-jane and the card not allocating anywhere near the amount of memory it could have (not running a display off the card).

[myagui]

Concerning the cards with more than 2GB, I think there is also some limitation in the driver layer (WDDM), 2.3GB or thereabouts is the maximum that Windows can use with regular display drivers. FWIW, I know the maximum VRAM I've used in Linux is about 3.7GB per card.

Windows 10 might be finally addressing this though (eventually):
GPU virtual memory in WDDM 2.0

[Grim]

Concerning the cards with more than 2GB, I think there is also some limitation in the driver layer (WDDM), 2.3GB or thereabouts is the maximum that Windows can use with regular display drivers. FWIW, I know the maximum VRAM I've used in Linux is about 3.7GB per card.

Windows 10 might be finally addressing this though (eventually):
GPU virtual memory in WDDM 2.0

What is your hashrate in linux with scrypt-n?

[sp_]

Upgrading the driver to the latest version is a good idea. NVIDIA fixed some memory allocation bugs.
With the older drivers, ccminer sometimes got out of memory even if there where alot of free memory on the cards.
Restarting the miner helped, but the crashes seamed to be random.


Windows 7 seems to be bether in allocating big blocks of memory on the gpu. In windows 8 there is some kind of restriction.. I have problems with the cryptonight miner to run with optimal settings.


On windows:

Since all my builds are 32 bit (faster hashing) you might want to increase the allowed memory in the application from 2gb to 3gb with this command:

editbin /LARGEADDRESSAWARE ccminer.exe

I will modify the project file tonight so that this flag is set automaticly in all future builds.

[Slava_K]

OK. Then i download and test on win 10...

[bensam1231]

Upgrading the driver to the latest version is a good idea. NVIDIA fixed some memory allocation bugs.
With the older drivers, ccminer sometimes got out of memory even if there where alot of free memory on the cards.
Restarting the miner helped, but the crashes seamed to be random.


Windows 7 seems to be bether in allocating big blocks of memory on the gpu. In windows 8 there is some kind of restriction.. I have problems with the cryptonight miner to run with optimal settings.


On windows:

Since all my builds are 32 bit (faster hashing) you might want to increase the allowed memory in the application from 2gb to 3gb with this command:

editbin /LARGEADDRESSAWARE ccminer.exe

I will modify the project file tonight so that this flag is set automaticly in all future builds.

I don't recall Claymore running into this, you may want to talk to him. He might have a idea of how to get around it or figure it out, unless this is a Nvidia only issue.

[SS2006]

I used to be able to get over 10000 KH and intensity 21 on a GTX 980
the latest version crashed at intensity 21 and at 20 it can't go over 9000

[sp_]

I assume this is x11?

Try with -i 20.9 or -i 20.8 and -g 2

-g 2 (2 threads) will actually increase the intensity by 1 but split the memory allocation into 2 blocks instead of one big block of memory. In windows 8 it seems to be a problem to allocate big chunks of continious gpu memory.


What is the rate when you don't have a -i parameter?

[myagui]

Concerning the cards with more than 2GB, I think there is also some limitation in the driver layer (WDDM), 2.3GB or thereabouts is the maximum that Windows can use with regular display drivers. FWIW, I know the maximum VRAM I've used in Linux is about 3.7GB per card.

Windows 10 might be finally addressing this though (eventually):
GPU virtual memory in WDDM 2.0

What is your hashrate in linux with scrypt-n?

I've never tested Scrypt-N at all, ever, so I am not familiar with the right settings for this one. If I have a chance I'll run some autotune/benchmarking and let you know.

[antonio8]

Upgrading the driver to the latest version is a good idea. NVIDIA fixed some memory allocation bugs.
With the older drivers, ccminer sometimes got out of memory even if there where alot of free memory on the cards.
Restarting the miner helped, but the crashes seamed to be random.


Windows 7 seems to be bether in allocating big blocks of memory on the gpu. In windows 8 there is some kind of restriction.. I have problems with the cryptonight miner to run with optimal settings.


On windows:

Since all my builds are 32 bit (faster hashing) you might want to increase the allowed memory in the application from 2gb to 3gb with this command:

editbin /LARGEADDRESSAWARE ccminer.exe

I will modify the project file tonight so that this flag is set automaticly in all future builds.

This errors out.
Says that "editbin" is not recognized as an internal or external command

[sp_]

It's a part of visual studio 2013.

https://msdn.microsoft.com/en-us/library/xd3shwhf.aspx

Don't worry I will add it to the project file so you don't need to do it.

[tbearhere]

@tbearhere

Can you post your hashrate/card for Scrypt-Jane:16 on Vista? Would be funny to find that Vista matches the performance on Linux (or betters it?), while recent Windows versions all suck! 

What does Linux do with the 750ti ?

[myagui]

No clue tbearhere, I've only got the 980.
Will send you a PM for more off-topic followup.

[chrysophylax]

@tbearhere

Can you post your hashrate/card for Scrypt-Jane:16 on Vista? Would be funny to find that Vista matches the performance on Linux (or betters it?), while recent Windows versions all suck! 

What does Linux do with the 750ti ?

let me know what settings need to be run in order to test this - and i shall do a 15minutes test tomorrow for you if you like ...

it will be with the latest git clone ... on nicehash using sp's donation link ... fedora 20 x64 ...

#crysx

[sp_]

Submitted another speedup in quark. More is comming.
I will improve the bandwidth of the hashes since half of the buffers are not used. Then the intensity can be increased..

[scryptr]

Submitted another speedup in quark. More is comming.
I will improve the bandwidth of the hashes since half of the buffers are not used. Then the intensity can be increased..

UNUSED BUFFERS--

Do unused buffers have anything to do with the poor performance of GTX 960/970 relative to GTX 750ti on Lyra2?  Or is it the memory controller?  Or both?  My 4GB GTX 960 is slower than a 2GB 750ti mining Lyra2, and it just indicates improper coding to match the card's capacity.  Both cards are on a Win 7 x64 system.

DJM34, don't be afraid to chime in, you are the code master of Lyra2.       --scryptr