CCminer(SP-MOD) Modded GPU kernels.

[chrysophylax]

Guys, which values do you reach on quark with r.53?
I've got 6.3-6.4 with 750ti, and 15.8-16.2 with 970...

hi ...

these hashrates do not look like these are stock clocks ... unless its factory overclocked in a big way ...

gigabyte 750ti oc lp cards run at 5600-5700KH with the occasional 5800KH thrown in for good measure ...

no overclock - no firmware fiddling ... just straight out of the box and hashing cards ...

i would luv to get over 6MH stock for these cards ...

but then again - we did get the lower end cards ...

#crysx

[chrysophylax]

6600 kh on quark with 750ti here

see ...

is this stock clocks? ...

they are nice stats ...

#crysx

[Zels]

OC +200core clock / +250 memory

[chrysophylax]

OC +200core clock / +250 memory

yup - thats what i thought ...

we really need to have some sort of standard that we can compare against ...

non oc is the best way on a card by card basis ...

every card oc's differently and some cards can be pushed harder than others - with others again being tweaked with firmware and such ...

this means that the readings we give here are utterly useless to compare with ...

they are great readings to compare oc'ing with - and a table / list that would be created as a comparison would be even better ...

who has skills ( and time ) to do such a thing? ...

is there already a site that has a comparison ( and settings ) list? ...

would luv to see how ccminer-spmod compares to other with the same cards ...

#crysx

[rednoW]

Some sad things about ccminer.
I used to test my "legendary" oc'ed gtx750 in different algo's on nicehash and yaamp.
Let's talk only about quark and nicehash as they are still look good ))

My ccminer shown hashrate was ~6000khs. But in nicehash on long run average hashrate was smoothly fluctuating between only 5400 and 5900. Very seldom it goes up to 6000 and very soon drops to ~5500 again. I talk about Average speed nicehash, not about Speed accepted.
I thought it's ok due to some network losses or something like that.

But now i temporary grab 2 radeon 7950 and put it instead of my gtx750 in the same machine. With optimized ccminer 5.1.1 (russian, available from cryptomining blog) I see about 21-21.5mhs in sgminer window. I use xintensity 512 that is known to provide good hashrate on pool ...
And nicehash now shows Average speed 21.5-22.3mhs. This is even better then miner shows!

So this situation is not in favor of nvidia and sp_'s ccminer. I know that nvidia and sp_'s ccminer are better in terms of perf per watt ... but this amd cards are so dirt cheap now ... and give nice absolute performance with more "honest" hashrate in miner ...

[Fuzzbawls]

OC +200core clock / +250 memory

yup - thats what i thought ...

we really need to have some sort of standard that we can compare against ...

non oc is the best way on a card by card basis ...

every card oc's differently and some cards can be pushed harder than others - with others again being tweaked with firmware and such ...

this means that the readings we give here are utterly useless to compare with ...

they are great readings to compare oc'ing with - and a table / list that would be created as a comparison would be even better ...

who has skills ( and time ) to do such a thing? ...

is there already a site that has a comparison ( and settings ) list? ...

would luv to see how ccminer-spmod compares to other with the same cards ...

#crysx

Would LOVE to see something like this! I know there have been various spreadsheets in the past, but looking at them now is only remotely useful with all the optimizations that have been made.

I usually get above 6000kh/s mining quark on my EVGA 750 Ti SC, its factory OCd and i've pushed it a little further too.

[sp_]

@SP
Release 53 x11 I am getting around a 5k increase (3180K on the upper end)..   Haven't changed intensity from version 52. 
I'm mining StartCoin on suchpool and it is showing my hash @4.06M w/v53. v52 it showed my hash has 2.6M.
Pool readings were taken after running 1 hr each time. 
4.06M has been holding steady for 3 hrs now, which is ~800K faster than the card is showing on confirms.
Cuda 7.0 and latest drivers (353.06).

Cuda 7.0 doesn't work/much slower. Use cuda 6.5 if you compile yourself.

[djm34]

Some sad things about ccminer.
I used to test my "legendary" oc'ed gtx750 in different algo's on nicehash and yaamp.
Let's talk only about quark and nicehash as they are still look good ))

My ccminer shown hashrate was ~6000khs. But in nicehash on long run average hashrate was smoothly fluctuating between only 5400 and 5900. Very seldom it goes up to 6000 and very soon drops to ~5500 again. I talk about Average speed nicehash, not about Speed accepted.
I thought it's ok due to some network losses or something like that.

But now i temporary grab 2 radeon 7950 and put it instead of my gtx750 in the same machine. With optimized ccminer 5.1.1 (russian, available from cryptomining blog) I see about 21-21.5mhs in sgminer window. I use xintensity 512 that is known to provide good hashrate on pool ...
And nicehash now shows Average speed 21.5-22.3mhs. This is even better then miner shows!

So this situation is not in favor of nvidia and sp_'s ccminer. I know that nvidia and sp_'s ccminer are better in terms of perf per watt ... but this amd cards are so dirt cheap now ... and give nice absolute performance with more "honest" hashrate in miner ...

several reasons:
* you set a too high diff or something like that...
* sgminer use a buffer to accumulate block header while solving others so it has always work on hand, and that's a big advantage on pool. It also send different blockheader to the different cards while in ccminer a new blockheader is obtained only after one has been solved and all cards work on the same block.
* sgminer 5 was developped by nicehash, so obviously they tuned their system on that...

[sp_]

Cuda 7.0 doesn't work/much slower. Use cuda 6.5 if you compile yourself.

I did some tests. Some of the algorithms are working with cuda 7.0. Like groestl, but the speed is 10% slower.
ALL aes implementations are spilling registers. 'Echo has 2048bytes spill stores with 64 registers in 7.0 and 160 spill bytes in 6.5
Shabal has 0 spill stores with cuda 6.5, but 500bytes in cuda 7.0. etc

One surprise in the sha256 in x17 is only using 56 registers with 0 spills in cuda 7.0 while it spills memory in 6.5...

[chrysophylax]

OC +200core clock / +250 memory

yup - thats what i thought ...

we really need to have some sort of standard that we can compare against ...

non oc is the best way on a card by card basis ...

every card oc's differently and some cards can be pushed harder than others - with others again being tweaked with firmware and such ...

this means that the readings we give here are utterly useless to compare with ...

they are great readings to compare oc'ing with - and a table / list that would be created as a comparison would be even better ...

who has skills ( and time ) to do such a thing? ...

is there already a site that has a comparison ( and settings ) list? ...

would luv to see how ccminer-spmod compares to other with the same cards ...

#crysx

I think a standard is good, but stock clocks probably shouldn't be it. Pure stock clocks are usually so bad, most people don't mine at them - hurts efficiency. A slight OC that pretty much all cards can reach makes sense, though.

shouldnt factory clocks BE the standard - even for a 25 minute test? ...

a list for stock ( standard rates ) and a list for oc ( using the same card and what has been done to the card ) ...

whether its is efficient or inefficient - a standard that the tweaks and overclocks ( with the methods of how it was done ) that can be compared to - is what i believe should be the 'norm' for comparison ...

for example - you buy a car that is 'factory stock' and test it ... then you tweak and tune and improve then test - and THEN compare the results ...

i believe it should be no different to gpu's ... within reason of course ...

no use trying to compare a liquid nitrogen cooled gpu with no way of duplicating it for the home user ...

as nice as that would be to see ...

#crysx

[rednoW]

several reasons:
* you set a too high diff or something like that...
* sgminer use a buffer to accumulate block header while solving others so it has always work on hand, and that's a big advantage on pool. It also send different blockheader to the different cards while in ccminer a new blockheader is obtained only after one has been solved and all cards work on the same block.
* sgminer 5 was developped by nicehash, so obviously they tuned their system on that...

I see. Concerning first * I didn't play with diff neither in ccminer nor in password provided to Nicehash. All is by default

[myagui]

FWIW, for me, mining quark, nicehash reports exactly the same hashrate as ccminer does locally (using any recent release on windows, 980s).
I've always kept an eye on the 24h averages, and they neatly match what ccminer reports, sometimes the pool reports a little higher, but just a negligible difference.

The instant hashrate (or small windows like 5 minutes) is rather pointless.
Going by instant hashrate, I've mined as slow as 10MH/s, and as fast as 80MH/s, you know, variance...

Just my 0.02 BTC ...

[chrysophylax]

OC +200core clock / +250 memory

yup - thats what i thought ...

we really need to have some sort of standard that we can compare against ...

non oc is the best way on a card by card basis ...

every card oc's differently and some cards can be pushed harder than others - with others again being tweaked with firmware and such ...

this means that the readings we give here are utterly useless to compare with ...

they are great readings to compare oc'ing with - and a table / list that would be created as a comparison would be even better ...

who has skills ( and time ) to do such a thing? ...

is there already a site that has a comparison ( and settings ) list? ...

would luv to see how ccminer-spmod compares to other with the same cards ...

#crysx

I think a standard is good, but stock clocks probably shouldn't be it. Pure stock clocks are usually so bad, most people don't mine at them - hurts efficiency. A slight OC that pretty much all cards can reach makes sense, though.

shouldnt factory clocks BE the standard - even for a 25 minute test? ...

a list for stock ( standard rates ) and a list for oc ( using the same card and what has been done to the card ) ...

whether its is efficient or inefficient - a standard that the tweaks and overclocks ( with the methods of how it was done ) that can be compared to - is what i believe should be the 'norm' for comparison ...

for example - you buy a car that is 'factory stock' and test it ... then you tweak and tune and improve then test - and THEN compare the results ...

i believe it should be no different to gpu's ... within reason of course ...

no use trying to compare a liquid nitrogen cooled gpu with no way of duplicating it for the home user ...

as nice as that would be to see ...

#crysx

The goals of the GPU manufacturer and the miner are very different. The GPU manufacturer wants to maximize the number of cards that pass the tests, so they set the standards low. Any miner caring about efficiency will OC somewhat if they can - we're not comparing mining usage to regular gaming usage, we're trying to compare a reasonable baseline vs improvements or better OCs.

With cars, whether you buy it to sit in traffic, or to race, it's still driving. Mining and gaming are two very different types of things - it's not just doing the same thing faster - which, I think, makes it reasonable to have a baseline more suited to mining (that almost all cards can still do) instead of one that was chosen by the manufacturer, who has entirely different goals in mind.

makes sense ...

so how would a baseline be agreed upon for comparison? ...

there really is no standards to base a lot of this off ( omg - here we go with the extranonce standards issue again ) ...

im curious ...

#crysx

[CapnBDL]

OC +200core clock / +250 memory

yup - thats what i thought ...

we really need to have some sort of standard that we can compare against ...

non oc is the best way on a card by card basis ...

every card oc's differently and some cards can be pushed harder than others - with others again being tweaked with firmware and such ...

this means that the readings we give here are utterly useless to compare with ...

they are great readings to compare oc'ing with - and a table / list that would be created as a comparison would be even better ...

who has skills ( and time ) to do such a thing? ...

is there already a site that has a comparison ( and settings ) list? ...

would luv to see how ccminer-spmod compares to other with the same cards ...

#crysx

I think a standard is good, but stock clocks probably shouldn't be it. Pure stock clocks are usually so bad, most people don't mine at them - hurts efficiency. A slight OC that pretty much all cards can reach makes sense, though.

shouldnt factory clocks BE the standard - even for a 25 minute test? ...

a list for stock ( standard rates ) and a list for oc ( using the same card and what has been done to the card ) ...

whether its is efficient or inefficient - a standard that the tweaks and overclocks ( with the methods of how it was done ) that can be compared to - is what i believe should be the 'norm' for comparison ...

for example - you buy a car that is 'factory stock' and test it ... then you tweak and tune and improve then test - and THEN compare the results ...

i believe it should be no different to gpu's ... within reason of course ...

no use trying to compare a liquid nitrogen cooled gpu with no way of duplicating it for the home user ...

as nice as that would be to see ...

#crysx

The goals of the GPU manufacturer and the miner are very different. The GPU manufacturer wants to maximize the number of cards that pass the tests, so they set the standards low. Any miner caring about efficiency will OC somewhat if they can - we're not comparing mining usage to regular gaming usage, we're trying to compare a reasonable baseline vs improvements or better OCs.

With cars, whether you buy it to sit in traffic, or to race, it's still driving. Mining and gaming are two very different types of things - it's not just doing the same thing faster - which, I think, makes it reasonable to have a baseline more suited to mining (that almost all cards can still do) instead of one that was chosen by the manufacturer, who has entirely different goals in mind.

makes sense ...

so how would a baseline be agreed upon for comparison? ...

there really is no standards to base a lot of this off ( omg - here we go with the extranonce standards issue again ) ...

im curious ...

#crysx

There used to be a spreadsheet out there that users could upload results, etc, too...

[chrysophylax]

FWIW, for me, mining quark, nicehash reports exactly the same hashrate as ccminer does locally (using any recent release on windows, 980s).
I've always kept an eye on the 24h averages, and they neatly match what ccminer reports, sometimes the pool reports a little higher, but just a negligible difference.

The instant hashrate (or small windows like 5 minutes) is rather pointless.
Going by instant hashrate, I've mined as slow as 10MH/s, and as fast as 80MH/s, you know, variance...

Just my 0.02 BTC ...

we find similar ...

though at time - the variance is quite large - as you stated ...

mining quark on nicehash us stratum ( westahsh ) for the last few hours - and this is proving to be true ...

besides - dont the pools 'average' the hashrate by the collective shares submitted by the miner? ...

if that is so - then it wont ever really be accurate ... isnt this the case? ...

#crysx

[chrysophylax]

OC +200core clock / +250 memory

yup - thats what i thought ...

we really need to have some sort of standard that we can compare against ...

non oc is the best way on a card by card basis ...

every card oc's differently and some cards can be pushed harder than others - with others again being tweaked with firmware and such ...

this means that the readings we give here are utterly useless to compare with ...

they are great readings to compare oc'ing with - and a table / list that would be created as a comparison would be even better ...

who has skills ( and time ) to do such a thing? ...

is there already a site that has a comparison ( and settings ) list? ...

would luv to see how ccminer-spmod compares to other with the same cards ...

#crysx

I think a standard is good, but stock clocks probably shouldn't be it. Pure stock clocks are usually so bad, most people don't mine at them - hurts efficiency. A slight OC that pretty much all cards can reach makes sense, though.

shouldnt factory clocks BE the standard - even for a 25 minute test? ...

a list for stock ( standard rates ) and a list for oc ( using the same card and what has been done to the card ) ...

whether its is efficient or inefficient - a standard that the tweaks and overclocks ( with the methods of how it was done ) that can be compared to - is what i believe should be the 'norm' for comparison ...

for example - you buy a car that is 'factory stock' and test it ... then you tweak and tune and improve then test - and THEN compare the results ...

i believe it should be no different to gpu's ... within reason of course ...

no use trying to compare a liquid nitrogen cooled gpu with no way of duplicating it for the home user ...

as nice as that would be to see ...

#crysx

The goals of the GPU manufacturer and the miner are very different. The GPU manufacturer wants to maximize the number of cards that pass the tests, so they set the standards low. Any miner caring about efficiency will OC somewhat if they can - we're not comparing mining usage to regular gaming usage, we're trying to compare a reasonable baseline vs improvements or better OCs.

With cars, whether you buy it to sit in traffic, or to race, it's still driving. Mining and gaming are two very different types of things - it's not just doing the same thing faster - which, I think, makes it reasonable to have a baseline more suited to mining (that almost all cards can still do) instead of one that was chosen by the manufacturer, who has entirely different goals in mind.

makes sense ...

so how would a baseline be agreed upon for comparison? ...

there really is no standards to base a lot of this off ( omg - here we go with the extranonce standards issue again ) ...

im curious ...

#crysx

There used to be a spreadsheet out there that users could upload results, etc, too...

you know - i do vaguely remember seeing that - now that you mention it ...

but where? ... and how accurate is it? ...

#crysx

[Angora]

[
[/quote]

you know - i do vaguely remember seeing that - now that you mention it ...

but where? ... and how accurate is it? ...

#crysx
[/quote]

Only 1 I've seen is for litecoin.      https://litecoin.info/Mining_hardware_comparison
We can use it as a guide to making one better suited to what we want.   

[CapnBDL]

I think a 5 hr. test would be enough, but... We are a bit off topic...a little. I'm on my tablet right now & am sure I have a bookmark for that stat sheet somewhere. I believe it's on my 'puter. I'll get back & look.

Later-

[myagui]

I think a 5 hr. test would be enough, but... We are a bit off topic...a little. I'm on my tablet right now & am sure I have a bookmark for that stat sheet somewhere. I believe it's on my 'puter. I'll get back & look.

Later-

Again, as I'm used to looking at the 24h graphs at nicehash, I can tell you that the reported average can easily go up or down by 5% or more, in a 5h window. This stands to produce a potential reading variance as high as 10% overall, or more. I'd put my finger at the 12h mark for the minimum period of reliable reporting, and ideally, the full 24h time frame.

On a slightly related note: people that benchmark often, and especially on the 970s & 980s, should take into account the temperature of their cards when trying to benchmark. From a cold start, my 980s always hash at their fastest for a couple of minutes, until they settle on their usual (hot) hashing temperature. If starting from a warm/hot state, then the output is more consistent throughout.
Anyone mining in the north pole or using liquid cooling will probably not care though...

[djm34]

several reasons:
* you set a too high diff or something like that...
* sgminer use a buffer to accumulate block header while solving others so it has always work on hand, and that's a big advantage on pool. It also send different blockheader to the different cards while in ccminer a new blockheader is obtained only after one has been solved and all cards work on the same block.
* sgminer 5 was developped by nicehash, so obviously they tuned their system on that...

I see. Concerning first * I didn't play with diff neither in ccminer nor in password provided to Nicehash. All is by default

I don't see how sgminer could cause a speedup by simply queueing work. I'm considering having my Stratum implementation fill up a global work queue for that specific pool, and then let the mining threads pop work as they will (unless interrupted by a hard restart due to a new block on the network or whatever else that would cause cleanjobs to be set in the mining.notify.) Thing is, it doesn't matter how much work I create and queue for the threads - if there's a new block on the network, I'm flushing all of it, because it's all gotta go - no work based on headers I generated before will be accepted. So, if ccminer just parses mining.notify, generates work, and passes it to all its running threads while restarting them, it'll have pretty much the exact same slowdown I will.

I don't say there is a speedup in sgminer but rather a slow down in ccminer... the advantage of sgminer is that it doesn't have to wait for a new block once it solved one. It pick up the next in its queue and also each gpu uses a different block which is better than the current implementation in ccminer (based on cpuminer) where the block is divided among the card (kinda pointless when you have fast and slow card on the same system.