CCminer(SP-MOD) Modded GPU kernels.

[bensam1231]

7x970 - Lyra2 won't start (out of memory)
4GB system memory

7x 970 ?!  mobo are so expensive...
you need at least as much ram than vram here something like 28Gb to run that kind of system

Surely there has to be a workaround. I mean the memory/swap doesn't even seem to be allocated let alone used, not even for a second.
Something like initializing the cards one after the other instead of all at the same time or something? Or giving the cards different jobs instead of working together on one big job? I have no idea but I'm sure there's a way.

Also agree, the same thing was happening to me with Neoscrypt and just had to throw more memory at it even though system memory basically isn't used at all.

If it just uses it to 'load' into the vram on the miner, a asynchronous load should help with it (load each card into memory, then into vram one at a time). Right now though I don't really see any indication of memory usage on the system.

[sp_]

7x970 - Lyra2 won't start (out of memory)
4GB system memory

-Upgrade to the latest NVidia drivers (22-jun-2015)
-Add 16GB virtual ram

NVIDIA fixed a memory allocation bug in their latest driver.

If it still doesn't work. reduce the intensity

f.eks -i 17

[sp_]

Its wrote on the main page, Yiimp is not an "autotrade" platform... So like others pools you mine the currency you want with the -right- currency address. I dont want to pay in VTC (or BTC) the whole china which is using SHA farms
The pool is working and pay what is mined... I don't want a second exchange full time job Consider the fees as a donation for the new algos... Some are set very high because we are doing "private" tests... you can still mine on those but its made to reduce "anonymous" users...

http://yiimp.ccminer.org/

This pool looks promising. Can you please add

sharkcoin(quark),
Digibyte(skein)
Myriadcoin(skein)

[bensam1231]

7x970 - Lyra2 won't start (out of memory)
4GB system memory

-Upgrade to the latest NVidia drivers (22-jun-2015)
-Add 16GB virtual ram

NVIDIA fixed a memory allocation bug in their latest driver.

If it still doesn't work. reduce the intensity

f.eks -i 17

I understand the whole throw more memory at it thing or reduce intensity, we basically went over the same thing when I was having Neo problems due to out of memory. Is there any reason the system doesn't actually use any memory and it's still getting these messages? The memory usage goes up slightly, but if you look at Resource Monitor, the system is barely using any of the available memory... Pagefile or Hardware.

[sp_]

SP_ RELEASE dot 54--
SP_'s reduction of register use in Lyra2 appears to have reduced some of the memory requirement.    I have been able to increase my intensity setting from my old standard of "-i 16.5" to higher values and see hash rate improvement, but the setting varies per machine.  My initial results, all for Lyra2:
  GTX 750ti FTW - 1080-1100kh/s per card  (Linux)
  GTX 750ti SC - 1140-1150kh/s per card (Win 8 )
  GTX 960 2GB SSC - 1220-1240kh/s per card (Win 7)
  GTX 960 4GB FTW - 1220-1240kh/s per card (Win 8 )
  GTX 970 4GB FTW+ - 2Mh/s per card (Linux)
The windows machines allow for easy software overclocking.  I still need to learn the command line API flags for Linux, and probably need to re-install Linux with the latest drivers for proper use.
If I move to CUDA Toolkit 7.5, will the SP_ releases still compile on Linux?       --scryptr

Since I use half the threads per block compared to djm34's version, you can add 1 to the maxintensity in the old miner and is still runs on the 750ti.
But running it without the intensity parameter should give a performance increase as well. Crypto mining blog messured it to be +150KHASH on the gtx 980 with the default intensity.

Since I am not a registred cuda developer, I cannot download cuda 7.5 and test.

[Epsylon3]

The windows machines allow for easy software overclocking.  I still need to learn the command line API flags for Linux, and probably need to re-install Linux with the latest drivers for proper use.

If I move to CUDA Toolkit 7.5, will the SP_ releases still compile on Linux?       --scryptr

Yes it will compile but most of the sp "fine tuning" is made "at the register" ... a kernel use a given number of registers which can diffear against the platform (os and 32/x64) and also the gpu sm.
If you change the OS or the sdk, some kernels will require to be retuned to fit a certain number of registers (one more reg. can reduce a lot the overall speed). I made this work for linux with nvprof because its faster to do and in general also benefits on windows...

The linux driver in the cuda 7.5 RC is the 352.07 and is older than the one i recommend (352.21) which have the power limit features. Else, you can install it easily on all distributions with the .run (tested ubuntu 14, debian 7, slackware and fedora 22) and both can be installed at once.

Regarding the "overclocking" functions, i think nvidia made a step, but didnt really finish the implementation... Power limit values seems to work, application clocks not sure except it change the pstate to P0...

http://yiimp.ccminer.org/

This pool looks promising. Can you please add

sharkcoin(quark),
Digibyte(skein)
Myriadcoin(skein)

Actually working on a proper way to mine only the coin set by your address, its why there is only one coin per algo for the moment...

[sp_]

The compiler in the cuda 7 produces shitty code. all the AES algos got increased registercount and the program is spilling memory, performance is lost, and the hash is broken.

Same on AMD (Omega drivers) performance is lost compared to the 14.6, 14.7 drivers.

[djm34]

7x970 - Lyra2 won't start (out of memory)
4GB system memory

7x 970 ?!   mobo are so expensive...
you need at least as much ram than vram here something like 28Gb to run that kind of system

Surely there has to be a workaround. I mean the memory/swap doesn't even seem to be allocated let alone used, not even for a second.
Something like initializing the cards one after the other instead of all at the same time or something? Or giving the cards different jobs instead of working together on one big job? I have no idea but I'm sure there's a way.

Also agree, the same thing was happening to me with Neoscrypt and just had to throw more memory at it even though system memory basically isn't used at all.

If it just uses it to 'load' into the vram on the miner, a asynchronous load should help with it (load each card into memory, then into vram one at a time). Right now though I don't really see any indication of memory usage on the system.

if you open msi AB and watch both ram and pagefile graphics, you'll see it gets allocated (more on the pagefile than on the memory) so may-be trying to increase pagefile could work.
There isn't really a work around on the code side, global memory variables have to be allocated from the host and cudamalloc works in mysterious way...)

[djm34]

The compiler in the cuda 7 produces shitty code. all the AES algos got increased register count and the program is spilling memory, performance is lost, and the hash is broken.

Same on AMD (Omega drivers) performance is lost compared to the 14.6, 14.7 drivers.

did you get an access violation in cuda_hefty ? How did you solved it ?  ;

cuda 7.5 gives however some performance boost in lyra 

The main problem is that it will be difficult to stay on some older version of cuda forever...

[sp_]

X11 with cuda7 is 10% slower. How fast is it with cuda 7,5?

[Epsylon3]

slower than 6.5 on windows

[pallas]

Same on AMD (Omega drivers) performance is lost compared to the 14.6, 14.7 drivers.

Depends on tha hash and on the card chip. Hawaii groestl got a 25% percent boost on 14.12 and whirlpoolx +10% on 15.3, for example.

[sp_]

http://hashpower.co/  (yaamp clone) is currently paying 0.7BTC/GHASH for quark. Have anyone tried this pool?

[pallas]

http://hashpower.co/  (yaamp clone) is currently paying 0.7BTC/GHASH for quark. Have anyone tried this pool?

I did, and didn't get my payments because they used a too low transaction fee.
Furthermore, payments have been stopped AFAICS.

[sp_]

http://hashpower.co/  (yaamp clone) is currently paying 0.7BTC/GHASH for quark. Have anyone tried this pool?

I did, and didn't get my payments because they used a too low transaction fee.
Furthermore, payments have been stopped AFAICS.

I will try it out, with payouts in DASH

Neoscrypt 20BTC/GHASH

[bensam1231]

7x970 - Lyra2 won't start (out of memory)
4GB system memory

7x 970 ?!   mobo are so expensive...
you need at least as much ram than vram here something like 28Gb to run that kind of system

Surely there has to be a workaround. I mean the memory/swap doesn't even seem to be allocated let alone used, not even for a second.
Something like initializing the cards one after the other instead of all at the same time or something? Or giving the cards different jobs instead of working together on one big job? I have no idea but I'm sure there's a way.

Also agree, the same thing was happening to me with Neoscrypt and just had to throw more memory at it even though system memory basically isn't used at all.

If it just uses it to 'load' into the vram on the miner, a asynchronous load should help with it (load each card into memory, then into vram one at a time). Right now though I don't really see any indication of memory usage on the system.

if you open msi AB and watch both ram and pagefile graphics, you'll see it gets allocated (more on the pagefile than on the memory) so may-be trying to increase pagefile could work.
There isn't really a work around on the code side, global memory variables have to be allocated from the host and cudamalloc works in mysterious way...)

So loading one card at a time, waiting for memory allocation, then loading another wouldn't help fix this? Do you guys already do this? The memory usage is increased, but there isn't anything indicating the system is anywhere close to out of memory, so when this happens, I would speculate it being a 'peak' allocation which happens right at the beginning of the mine where a lot of things are loaded into memory and instantly loaded into vram, but that instant is enough to push memory usage over the top.

Maybe I'm mistaken about that. It could just be 'holding' memory too after it moves it from system memory to vram even though it wont really ever use that much memory again.

http://hashpower.co/  (yaamp clone) is currently paying 0.7BTC/GHASH for quark. Have anyone tried this pool?

I did, and didn't get my payments because they used a too low transaction fee.
Furthermore, payments have been stopped AFAICS.

I will try it out, with payouts in DASH

Neoscrypt 20BTC/GHASH

Looks as though that's about what Nicehash is currently paying.

Also keep in mind because it's such a small pool this could be finders 'luck' and not a expected payout. Basically the pool got lucky finding blocks and it's not big enough to give you a realistic representation of payouts.

[sp_]

Submittet a bugfix and a speedup in quark.

The gtx windforce 970 is now peaking at 16130 on standard clocks.  (up from 15800)

[dga]

Submittet a bugfix and a speedup in quark.

The gtx windforce 970 is now peaking at 16130 on standard clocks.  (up from 15800)

Note that release 54 has a small bug in the hash that will report lower rates on the pool.

Just fyi, gtx 980:

Code:

[2015-07-09 15:24:05] GPU #0: result for nonce $0353D188 does not validate on CPU!
[2015-07-09 15:24:09] GPU #1: result for nonce $8BFDF7B0 does not validate on CPU!

compiling with cuda 7.0 on ubuntu for 980.

Does work under 6.5 on a 750ti.  Not sure if it's card or cuda version.

[sp_]

If you want a working miner for cuda7.0 this is the correct branch:

https://github.com/tpruvot/ccminer

I tried to compile my fork for the cuda 7.0 and started modding a bit. But the compiler wasn't good enough. Not worth the effort..
Around 10% drop in hashrate in all algos..

If you have time please run tvprovot's cuda 7.0 version of quark and compare the hashrate with my 6.5 release 54-git and post your findings.

[flipclip]

How do you do checkouts then?  I'm used to the command line and using the sha for checkouts, so that is why I am wondering.

COMMAND LINE--

I use the command line, and refer to the commit number when posting about performance.  The sha will verify checksum, and is very precise for that purpose.  Commit numbers are sequential.

The line, "git clone https://github.com/sp-hash/ccminer", should clone the latest commit.  If I am wrong, please tell me!

--scryptr

Yes that is the correct command to get the latest commit.  I was just wondering if you used some type of command to get a previous commit via "COMMIT #".  An example would be someone says "COMMIT #820 is doing faster lyra still" (completely made up example by the way), but you are already on COMMIT #843, so you use a git command to go back to COMMIT #820 to try compiling from that point in the past.  I've always used "git checkout sha" to go "back in time" (hence the reason I always find the sha information more informative then "COMMIT #") but I was thinking maybe I was missing a command.