CCminer(SP-MOD) Modded GPU kernels.

[djm34]

The source got released, but he has done a private neoscrypt kernal that does 750khs/s on the 980 up from 630.

My free opensource sp-modded version is only around 7% faster than 630. around 680khash

now doing 820kh/s on the 980
               496kh/s on the 780
               196kh/s on the 750ti (my 750ti was just doing 165kh/s before...)

[tbearhere]

The 980 ti
http://www.gigabyte.com/products/product-page.aspx?pid=5208#ov

Would there be a difference if the card where 6gb and not 4gb ram on hashrate ?
thx

[tbearhere]

The source got released, but he has done a private neoscrypt kernal that does 750khs/s on the 980 up from 630.

My free opensource sp-modded version is only around 7% faster than 630. around 680khash

now doing 820kh/s on the 980
               496kh/s on the 780
               196kh/s on the 750ti (my 750ti was just doing 165kh/s before...)

Good improvement.
How do I get the updated miner djm?

[bathrobehero]

The 980 ti
http://www.gigabyte.com/products/product-page.aspx?pid=5208#ov

Would there be a difference if the card where 6gb and not 4gb ram on hashrate ?
thx

That's a non-Ti.

This is a Ti which have 6GB memory: http://www.gigabyte.com/products/product-page.aspx?pid=5472#ov

[djm34]

The source got released, but he has done a private neoscrypt kernal that does 750khs/s on the 980 up from 630.

My free opensource sp-modded version is only around 7% faster than 630. around 680khash

now doing 820kh/s on the 980
               496kh/s on the 780
               196kh/s on the 750ti (my 750ti was just doing 165kh/s before...)

Good improvement.
How do I get the updated miner djm?

you pray 

[djm34]

The 980 ti
http://www.gigabyte.com/products/product-page.aspx?pid=5208#ov

Would there be a difference if the card where 6gb and not 4gb ram on hashrate ?
thx

to answer the question, just look at how much memory you use on your card, if you are close to the 4gb limit, might help if not well it won't...
(the amount of vram isn't a bottleneck in any algo, rather memory bandwidth)
actually memory bandwidth is higher, so that's part might improved a bit things up  

anyhow, probably better to wait, there are some rumors which say that the price will go down a bit...

[myagui]

[...] the amount of vram isn't a bottleneck in any algo, rather memory bandwidth [...]

With Scrypt-Jane at high n-factors (and possibly with Scrypt-N too, that one I have no experience with), greater amounts of VRAM make a huge difference. Sidenote: only for Linux, since Windows sux and can't ever fully utilize the VRAM.
I suppose these are somewhat the exception, since most other algos don't really care for memory quantity at all.

[bensam1231]

The source got released, but he has done a private neoscrypt kernal that does 750khs/s on the 980 up from 630.

My free opensource sp-modded version is only around 7% faster than 630. around 680khash

now doing 820kh/s on the 980
               496kh/s on the 780
               196kh/s on the 750ti (my 750ti was just doing 165kh/s before...)

What is the hash/watt stock clocks? Really should be included with any advertisement of hashrate at this point.

[...] the amount of vram isn't a bottleneck in any algo, rather memory bandwidth [...]

With Scrypt-Jane at high n-factors (and possibly with Scrypt-N too, that one I have no experience with), greater amounts of VRAM make a huge difference. Sidenote: only for Linux, since Windows sux and can't ever fully utilize the VRAM.
I suppose these are somewhat the exception, since most other algos don't really care for memory quantity at all.

Thought monero also uses a lot of memory.

[myagui]

Thought monero also uses a lot of memory.

Relatively speaking, I think it does, but with Monero, I haven't seen any difference between similar 2GB and 4GB cards. So while it uses a lot of VRAM, it cares more about memory bandwidth, less about absolute memory quantity (my experience is only with 2GB up).

Obviously, the developers will know much better, but my basic understanding is that with Monero, you just need enough VRAM to fit a certain/large scratchpad. Beyond that, more memory does not help in any meaningful way. Faster memory does help.
There might be coding optimizations/tradeoffs to be made, much like larger -L values do for Scrypt-Jane & Scrypt-N (not storing the full scratchpad and computing on the fly as needed for missing values).

Hope the description makes some sense. If not, may the bashing begin 

[tbearhere]

The 980 ti
http://www.gigabyte.com/products/product-page.aspx?pid=5208#ov

Would there be a difference if the card where 6gb and not 4gb ram on hashrate ?
thx

That's a non-Ti.

This is a Ti which have 6GB memory: http://www.gigabyte.com/products/product-page.aspx?pid=5472#ov

Yes thx  I thought I was looking at a 980 ti with 4GB and that's why I got confused.

[tbearhere]

The source got released, but he has done a private neoscrypt kernal that does 750khs/s on the 980 up from 630.

My free opensource sp-modded version is only around 7% faster than 630. around 680khash

now doing 820kh/s on the 980
               496kh/s on the 780
               196kh/s on the 750ti (my 750ti was just doing 165kh/s before...)

Good improvement.
How do I get the updated miner djm?

you pray 

I contributed....do you want more beers. 

[tbearhere]

The 980 ti
http://www.gigabyte.com/products/product-page.aspx?pid=5208#ov

Would there be a difference if the card where 6gb and not 4gb ram on hashrate ?
thx

to answer the question, just look at how much memory you use on your card, if you are close to the 4gb limit, might help if not well it won't...
(the amount of vram isn't a bottleneck in any algo, rather memory bandwidth)
actually memory bandwidth is higher, so that's part might improved a bit things up  

anyhow, probably better to wait, there are some rumors which say that the price will go down a bit...

That's the problem ...I can't wait much longer.    But thx djm

[tbearhere]

[...] the amount of vram isn't a bottleneck in any algo, rather memory bandwidth [...]

With Scrypt-Jane at high n-factors (and possibly with Scrypt-N too, that one I have no experience with), greater amounts of VRAM make a huge difference. Sidenote: only for Linux, since Windows sux and can't ever fully utilize the VRAM.
I suppose these are somewhat the exception, since most other algos don't really care for memory quantity at all.

Yes the latest versions of windows does... but my vista does 3 x what windows 8.1 does on hashrate with Scrypt-Jane  16.

[myagui]

@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! 

[bathrobehero]

For scrypt based coins ccminer is set to only use 88% of the maximum video RAM in case of Windows and 98% in case of Linux by default:

Code:

device_backoff[i] = is_windows() ? 12 : 2;

It is there for safety so it won't crash for people who wish to mine on their main card.
One of the changes I always do before compiling is to change that value to 1 for Windows to allow it to use 99% of memory and if you don't have a monitor attached to your card and not running something that is using VRAM needlessly, it won't crash. Not even while autotuning.


I used that since cudaminer days so I have no idea about the hashrate difference anymore.


Now that we're entering the territory of cards with 3GB+ video memory it's worth noting that 32-bit ccminer can't utilize more than 3GB.
64-bit can utilize it all but 64-bit is also noticeably slower so for cards like the 970 it doesn't worth it but for the 980 Ti it might - for algos that scale with memory amount (not access) like Scrypt-Jane.
Afaik, 32-bit softwares can actually utilize more than 3GB video memory but only if the data is also stored in the system RAM which is generally not optimal for mining.

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

[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! 

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.

[Grim]

For scrypt based coins ccminer is set to only use 88% of the maximum video RAM in case of Windows and 98% in case of Linux by default:

Code:

device_backoff[i] = is_windows() ? 12 : 2;

It is there for safety so it won't crash for people who wish to mine on their main card.
One of the changes I always do before compiling is to change that value to 1 for Windows to allow it to use 99% of memory and if you don't have a monitor attached to your card and not running something that is using VRAM needlessly, it won't crash. Not even while autotuning.


I used that since cudaminer days so I have no idea about the hashrate difference anymore.


Now that we're entering the territory of cards with 3GB+ video memory it's worth noting that 32-bit ccminer can't utilize more than 3GB.
64-bit can utilize it all but 64-bit is also noticeably slower so for cards like the 970 it doesn't worth it but for the 980 Ti it might - for algos that scale with memory amount (not access) like Scrypt-Jane.
Afaik, 32-bit softwares can actually utilize more than 3GB video memory but only if the data is also stored in the system RAM which is generally not optimal for mining.

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

cudaminer allocates a maximum of 1.05 GB for scrypt-n (chacha) on my 750ti cards. If you use settings which allocate beyond that 1.05 GB the miner slows down considerably.

eg. I had a chance to try to mine with a 750ti 4gb !!! but any settings above 1.05 GB killed the performance.

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)

For team red yacminer is very neat because you can tweak/allocate every single megabyte of vram which lets you set a lower lookup gap and higher intensity.

[Grim]

@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!  

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

[tbearhere]

For scrypt based coins ccminer is set to only use 88% of the maximum video RAM in case of Windows and 98% in case of Linux by default:

Code:

device_backoff[i] = is_windows() ? 12 : 2;

It is there for safety so it won't crash for people who wish to mine on their main card.
One of the changes I always do before compiling is to change that value to 1 for Windows to allow it to use 99% of memory and if you don't have a monitor attached to your card and not running something that is using VRAM needlessly, it won't crash. Not even while autotuning.


I used that since cudaminer days so I have no idea about the hashrate difference anymore.


Now that we're entering the territory of cards with 3GB+ video memory it's worth noting that 32-bit ccminer can't utilize more than 3GB.
64-bit can utilize it all but 64-bit is also noticeably slower so for cards like the 970 it doesn't worth it but for the 980 Ti it might - for algos that scale with memory amount (not access) like Scrypt-Jane.
Afaik, 32-bit softwares can actually utilize more than 3GB video memory but only if the data is also stored in the system RAM which is generally not optimal for mining.

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

How does one apply that?

device_backoff = is_windows() ? 12 : 2;

thx

[tbearhere]

For scrypt based coins ccminer is set to only use 88% of the maximum video RAM in case of Windows and 98% in case of Linux by default:

Code:

device_backoff[i] = is_windows() ? 12 : 2;

It is there for safety so it won't crash for people who wish to mine on their main card.
One of the changes I always do before compiling is to change that value to 1 for Windows to allow it to use 99% of memory and if you don't have a monitor attached to your card and not running something that is using VRAM needlessly, it won't crash. Not even while autotuning.


I used that since cudaminer days so I have no idea about the hashrate difference anymore.


Now that we're entering the territory of cards with 3GB+ video memory it's worth noting that 32-bit ccminer can't utilize more than 3GB.
64-bit can utilize it all but 64-bit is also noticeably slower so for cards like the 970 it doesn't worth it but for the 980 Ti it might - for algos that scale with memory amount (not access) like Scrypt-Jane.
Afaik, 32-bit softwares can actually utilize more than 3GB video memory but only if the data is also stored in the system RAM which is generally not optimal for mining.

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

How does one apply that?

device_backoff = is_windows() ? 12 : 2;

thx

Format for the ternary operator is expression ? if_true : if_false;

Being a noob at this....I better leave this alone.
Thx  Wolf0  

EDIT: Just curious how one goes about this.