Author

Topic: CCminer(SP-MOD) Modded NVIDIA Maxwell / Pascal kernels. - page 1141. (Read 2347601 times)

legendary
Activity: 2716
Merit: 1094
Black Belt Developer
i don't remember now but, can this miner mine cryptonite?

no mate ... but sp has his private miner that is optimized for that algo ...

#crysx

What kind of hash for a GTX 750 ti? Will 660 ti work?

i cant exactly remember - but it was ok ...

from memory - tsiv's miner ( which sp optimized - and correct me if im wrong sp ) was close but didnt hash as much as sp's optimized miner ...

i need sleep now - so i could be way off the mark - but i can test it again for you tomorrow to get an exact hashrate if you like ...

#crysx

It is a cost-benefit thing. If there is enough extra hashpower to cover the cost relatively shortly, through mining XMR, then it would be worth the cost. Depends on the cost though.

I would be more happy with miners that took a percentage fee from the hashpower vs a one time charge for it. Sort of a willing botnet where the dev keeps getting paid to infinity and makes it worthwhile to keep developing and improving.

the problem with miner and a percentage fee is that, to be honest, a developer should code his own miner (not use a modified opensource one). It also moves the risk from the miner to the developer.
legendary
Activity: 1582
Merit: 1019
011110000110110101110010
i don't remember now but, can this miner mine cryptonite?

no mate ... but sp has his private miner that is optimized for that algo ...

#crysx

What kind of hash for a GTX 750 ti? Will 660 ti work?

i cant exactly remember - but it was ok ...

from memory - tsiv's miner ( which sp optimized - and correct me if im wrong sp ) was close but didnt hash as much as sp's optimized miner ...

i need sleep now - so i could be way off the mark - but i can test it again for you tomorrow to get an exact hashrate if you like ...

#crysx

It is a cost-benefit thing. If there is enough extra hashpower to cover the cost relatively shortly, through mining XMR, then it would be worth the cost. Depends on the cost though.

I would be more happy with miners that took a percentage fee from the hashpower vs a one time charge for it. Sort of a willing botnet where the dev keeps getting paid to infinity and makes it worthwhile to keep developing and improving.
legendary
Activity: 1154
Merit: 1001
cryptonite =! cryptonight  Wink
legendary
Activity: 2912
Merit: 1091
--- ChainWorks Industries ---
i don't remember now but, can this miner mine cryptonite?

no mate ... but sp has his private miner that is optimized for that algo ...

#crysx

What kind of hash for a GTX 750 ti? Will 660 ti work?

i cant exactly remember - but it was ok ...

from memory - tsiv's miner ( which sp optimized - and correct me if im wrong sp ) was close but didnt hash as much as sp's optimized miner ...

i need sleep now - so i could be way off the mark - but i can test it again for you tomorrow to get an exact hashrate if you like ...

#crysx
legendary
Activity: 1582
Merit: 1019
011110000110110101110010
i don't remember now but, can this miner mine cryptonite?

no mate ... but sp has his private miner that is optimized for that algo ...

#crysx

What kind of hash for a GTX 750 ti? Will 660 ti work?
legendary
Activity: 2912
Merit: 1091
--- ChainWorks Industries ---
i don't remember now but, can this miner mine cryptonite?

no mate ... but sp has his private miner that is optimized for that algo ...

#crysx
legendary
Activity: 3248
Merit: 1070
i don't remember now but, can this miner mine cryptonite?
legendary
Activity: 3164
Merit: 1003
sp  how much does electricity cost there?
legendary
Activity: 2912
Merit: 1091
--- ChainWorks Industries ---
sp could you ask them how much electricity cost?  plz  Wink
i too would be interested in shifting the farm there ... i mean - interested in electricity costs ... cough cough ...
Wink ...
#crysx

Today it was 38 degrees celcius outside. This is not a good place to mine. Smiley

here in adelaide it gets to 45C ...

so its not an issue mate - just got to know how to manipulate the airflow ...

#crysx
sr. member
Activity: 506
Merit: 252
So why is the 750ti performing best (hash per watt) with a rather low gddr5 bandwith? (cryptonight)
The only thing it offers is a huge level 2 cache (2mb).

sp_
legendary
Activity: 2926
Merit: 1087
Team Black developer
In spreadcoin it would be possible if the buffer was a bit smaller.
)
The bottleneck in the spreadcoin hash is a 200 000 table sha256 runthrough. (800kb)

The 980 has 96kb shared mem (level 1 cache)

So you would need 9 980 cards in a rig.

card 1 hash the first 96KB, passes the result to card2 wich calculates the next 96kb block, then card3 etc..

Each card will store the partial result in their sharedmem/level1 cache.
legendary
Activity: 2716
Merit: 1094
Black Belt Developer
If I recall correctly, cryptonight uses a 2MB buffer, which means you need MORE than 2MB of cache to fit it entirely and avoid using main gpu ram. Moreover, the miners must work on many cards so they probably make use of gpu ram anyway. Cache helps, but there will be a lot of misses. Thus, cache size and speed matters but so does global memory latency.

If your rig has 6 gpu's they have a total of 12 MB cache. instead of computing the complete hash each of the gpu's, it might be possible to spread the work.

not sure the internal state can be split... maybe someone with better knowledge on cryptonight, like wolf0, can comment.
sp_
legendary
Activity: 2926
Merit: 1087
Team Black developer
If I recall correctly, cryptonight uses a 2MB buffer, which means you need MORE than 2MB of cache to fit it entirely and avoid using main gpu ram. Moreover, the miners must work on many cards so they probably make use of gpu ram anyway. Cache helps, but there will be a lot of misses. Thus, cache size and speed matters but so does global memory latency.

If your rig has 6 gpu's they have a total of 12 MB cache. instead of computing the complete hash each of the gpu's, it might be possible to spread the work.
sp_
legendary
Activity: 2926
Merit: 1087
Team Black developer
i understand cuda 7 will have a few more optimized tweaks and stability for compiling ...
#crysx

cuda 7 is released now. I think I will compile version 45 with cuda 7. need to test if everything works though. The beta version of cuda 7 crashed my fork of ccminer.
sp_
legendary
Activity: 2926
Merit: 1087
Team Black developer
sp could you ask them how much electricity cost?  plz  Wink
i too would be interested in shifting the farm there ... i mean - interested in electricity costs ... cough cough ...
Wink ...
#crysx

Today it was 38 degrees celcius outside. This is not a good place to mine. Smiley
legendary
Activity: 2912
Merit: 1091
--- ChainWorks Industries ---
1 bottle of 40% thai whiskey costs 0,0088487 BTC
Suddenly, I want to go to Thailand. I'm afraid I might stay there though.

But thaiwhiskey is made from rice and not potatoes. Doesn't taste as good as what we are used to..
Prefer beer.

Today I ate a bigmac at McDonald's 0,01453715 BTC.

More expensive than 1 hour of thai-massage.

And in the tourist areas you can multiply the prices by 3-5
sp could you ask them how much electricity cost?  plz  Wink

i too would be interested in shifting the farm there ... i mean - interested in electricity costs ... cough cough ...

Wink ...

#crysx
legendary
Activity: 2912
Merit: 1091
--- ChainWorks Industries ---
Is there a specific version of Cuda I should have installed?
If you just want to use the binaries you only need the current driver.
If you want to compile it yourself you need the CUDA Toolkit 6.5.

i still compile with 6.0 ...

only due to the crash issues and 'cannot detect cuda devices' when compiled with 6.5 ...

i understand cuda 7 will have a few more optimized tweaks and stability for compiling ...

waiting for the fedora 21 repo to come out ...

#crysx
legendary
Activity: 2912
Merit: 1091
--- ChainWorks Industries ---
comparing v43 vs. v44 for my gtx750(non-ti) (windows binaries from github)
x15: 1966 vs 1976
x13: 2272 vs 2270
x11: 2886 vs 2883
quark: 5700 vs 5660
qubit: 4418 vs 4419 (with drops)
lyra: 882 vs 881
So, only x15 gain for me

Your junior 750 is beating some 750ti's on standard clocks.
I think it's time to implement a command line switch for the gpu-memspeed and the gpu-core clock. Some of the algorithms needs a good memoryspeed/corespeed factor, and some of the cards are not configured optimal. I had a quick look and found support in the NVIDIA api to adjust the clocks.. Will have a look when I get back from my vacation..

In the north east thailand:

1 hour of Thai-massage costs        0.012641 BTC
1 bottle of 40% thai whiskey costs 0,0088487 BTC
1 noudlesoup meal with meat and clean ice water in the resturant costs 0,0029495  BTC

nice Smiley ...

would be ideal if we could manipulate the clocks ( mem and core ) from the commandline ...

sounds very cheap over there ... and very nice Smiley ...

might even be able to buy property too with btc ... i wonder ...

Wink ...

#crysx

Linux command line to change clocks on my laptop, GTX850M, this is in my launch script.
optirun nvidia-settings -c :8 -a [gpu:0]/GPUGraphicsClockOffset[2]=80 -a [gpu:0]/GPUMemoryTransferRateOffset[2]=300

If you dont run optimus/bumblebee to switch on the second graphics card you dont need optirun, :8 is the display number, you need to have coolbits set in your config for it to work.


tanx ...

will need to try that out once i have some time ...

i run fedora 19 x64 on the work machines - so its changed a great deal from the days of fedoracore ...

will figure it out though Wink ...

btw - would anyone know the 'stable' oc settings for gigabyte 750ti oc lp cards? ...

tanx again ...

#crysx
legendary
Activity: 1154
Merit: 1001
What really matters in this case, is memory bandwidth. The GPU miner runs a shitload of threads, and 2MB fits a single thread. I don't think the GPUs would be doing much hashing if running in single thread mode in order to use the cache exclusively, as opposed to using the much slower GPU RAM and executing a very large number of concurrent threads. Grin

That is also why cryptonight CPU mining, you run X threads based on how much cache the CPU has. A CPU with 6MB cache, runs optimally with 3 threads, a CPU with 8MB cache, runs optimally with 4 threads, and so on.
legendary
Activity: 2716
Merit: 1094
Black Belt Developer
Regarding cryptonight:

for the ccminer cryptonight kernel it seems like the level 2 cache is a big deciding factor.

750ti   / 2mb l2 cache
960     / 1mb l2 cache
970     / 1.75mb l2 cache
980     / 2mb l2 cache
Titan X / 3mb l2 cache (anyone got one yet  Shocked )



But how come the amd cards are that fast with only 768kb for the 280x or 1mb for the 290x??
Is memory bandwidth and latency that much more important for this algo?

If I recall correctly, cryptonight uses a 2MB buffer, which means you need MORE than 2MB of cache to fit it entirely and avoid using main gpu ram. Moreover, the miners must work on many cards so they probably make use of gpu ram anyway. Cache helps, but there will be a lot of misses. Thus, cache size and speed matters but so does global memory latency.
Jump to: