[ANN] sgminer v5 - optimized X11/X13/NeoScrypt/Lyra2RE/etc. kernel-switch miner

[winmkx]

Has anybody done per-algorithm* comparison of CPU and GPU megahashes? And/or the wattage on GPU.

Like for example: BMW 100MHS on R9 280X and 20MHS on Core i7 8 threads.

Which of the SPH algorithms are the best for CPU?

* By algoritm here I mean the primitive algos, like BMW, skein, luffa, etc. Not the combined ones like Quark, X11, X15, etc

Best for CPU compared to GPU? None. You want a KDF, not a hash function.

No, not better on CPU. I'd say "less bad" on CPU than on GPU. Perhaps on some algorithms CPU is 100X worse, while on others it is 5X worse?

For example, CPUs have built in AES NI, so algorithms like Fugue should be not as lousy on CPU?

[1714797090]

1714797090

[1714797090]

1714797090

[winmkx]

OK, I did some measurements on CPU and GPU and here are the results below:

Detailed CPU results can be obtained from eBACS for multiple architectures, but I used my own quick code below on Core i7:

Code:

#include "stdafx.h"
#include <stdlib.h>
#include <stdint.h>
#include <string.h>
#include <stdio.h>
#include <time.h>

#include "sha3/sph_blake.h"
#include "sha3/sph_bmw.h"
#include "sha3/sph_groestl.h"
#include "sha3/sph_jh.h"
#include "sha3/sph_keccak.h"
#include "sha3/sph_skein.h"
#include "sha3/sph_luffa.h"
#include "sha3/sph_cubehash.h"
#include "sha3/sph_shavite.h"
#include "sha3/sph_simd.h"
#include "sha3/sph_echo.h"
#include "sha3/sph_hamsi.h"
#include "sha3/sph_fugue.h"
#include "sha3/sph_shabal.h"
#include "sha3/sph_whirlpool.h"

sph_blake512_context     ctx_blake;
sph_bmw512_context       ctx_bmw;
sph_groestl512_context   ctx_groestl;
sph_jh512_context        ctx_jh;
sph_keccak512_context    ctx_keccak;
sph_skein512_context     ctx_skein;
sph_luffa512_context     ctx_luffa;
sph_cubehash512_context  ctx_cubehash;
sph_shavite512_context   ctx_shavite;
sph_simd512_context      ctx_simd;
sph_echo512_context      ctx_echo;
sph_hamsi512_context     ctx_hamsi;
sph_fugue512_context     ctx_fugue;
sph_shabal512_context    ctx_shabal;
sph_whirlpool_context    ctx_whirlpool;

#define TEST_PRE(na) printf("%s: ", na); \
unsigned char hash[256]; \
for(int i=0; i<64; i++) hash[i] = i; \
int t = clock(); \
for(int i=0; i<1000000; i++) \
{

#define TEST_POST() } \
	printf ("\t\t%g seconds\n", ((float)(clock() - t)) / CLOCKS_PER_SEC); \

void main()
{
	{//blake512
		TEST_PRE("blake512");
		sph_blake512_init(&ctx_jh);
		sph_blake512(&ctx_jh, hash, 64);
		sph_blake512_close(&ctx_jh, hash);
		TEST_POST();
	}

	{//bmw512
		TEST_PRE("bmw512");
		sph_bmw512_init(&ctx_jh);
		sph_bmw512(&ctx_jh, hash, 64);
		sph_bmw512_close(&ctx_jh, hash);
		TEST_POST();
	}

	{//groestl512
		TEST_PRE("groestl512");
		sph_groestl512_init(&ctx_jh);
		sph_groestl512(&ctx_jh, hash, 64);
		sph_groestl512_close(&ctx_jh, hash);
		TEST_POST();
	}

	{//skein512
		TEST_PRE("skein512");
		sph_skein512_init(&ctx_jh);
		sph_skein512(&ctx_jh, hash, 64);
		sph_skein512_close(&ctx_jh, hash);
		TEST_POST();
	}

	{//jh512
		TEST_PRE("jh512");
		sph_jh512_init(&ctx_jh);
		sph_jh512(&ctx_jh, hash, 64);
		sph_jh512_close(&ctx_jh, hash);
		TEST_POST();
	}

	{//keccak512
		TEST_PRE("keccak512");
		sph_keccak512_init(&ctx_jh);
		sph_keccak512(&ctx_jh, hash, 64);
		sph_keccak512_close(&ctx_jh, hash);
		TEST_POST();
	}

	{//luffa512
		TEST_PRE("luffa512");
		sph_luffa512_init(&ctx_jh);
		sph_luffa512(&ctx_jh, hash, 64);
		sph_luffa512_close(&ctx_jh, hash);
		TEST_POST();
	}

	{//cubehash512
		TEST_PRE("cubehash512");
		sph_cubehash512_init(&ctx_jh);
		sph_cubehash512(&ctx_jh, hash, 64);
		sph_cubehash512_close(&ctx_jh, hash);
		TEST_POST();
	}

	{//shavite512
		TEST_PRE("shavite512");
		sph_shavite512_init(&ctx_jh);
		sph_shavite512(&ctx_jh, hash, 64);
		sph_shavite512_close(&ctx_jh, hash);
		TEST_POST();
	}

	{//simd512
		TEST_PRE("simd512");
		sph_simd512_init(&ctx_jh);
		sph_simd512(&ctx_jh, hash, 64);
		sph_simd512_close(&ctx_jh, hash);
		TEST_POST();
	}

	{//echo512
		TEST_PRE("echo512");
		sph_echo512_init(&ctx_jh);
		sph_echo512(&ctx_jh, hash, 64);
		sph_echo512_close(&ctx_jh, hash);
		TEST_POST();
	}

	{//hamsi512
		TEST_PRE("hamsi512");
		sph_hamsi512_init(&ctx_jh);
		sph_hamsi512(&ctx_jh, hash, 64);
		sph_hamsi512_close(&ctx_jh, hash);
		TEST_POST();
	}

	{//fugue512
		TEST_PRE("fugue512");
		sph_fugue512_init(&ctx_jh);
		sph_fugue512(&ctx_jh, hash, 64);
		sph_fugue512_close(&ctx_jh, hash);
		TEST_POST();
	}

	{//shabal512
		TEST_PRE("shabal512");
		sph_shabal512_init(&ctx_jh);
		sph_shabal512(&ctx_jh, hash, 64);
		sph_shabal512_close(&ctx_jh, hash);
		TEST_POST();
	}

	{//whirlpool
		TEST_PRE("whirlpool");
		sph_whirlpool_init(&ctx_jh);
		sph_whirlpool(&ctx_jh, hash, 64);
		sph_whirlpool_close(&ctx_jh, hash);
		TEST_POST();
	}
}

All results are scaled to be relative to the fastest hash function, which is shabal.

And here's the graph:
https://i.imgur.com/ySLE1xK.png


GPU results where obtained on R9 280X with smginer 5.1 and driver 13.X using AMD CodeXL and scaled to be relative to the fastest function shabal:
https://i.imgur.com/k6lif5s.png


Notice that CPU speed difference is 1X - 8X, while GPU speed difference is 1X - 45X.


PS: Note, you cannot compare values from CPU graph and GPU graph directly. Only algorithm's relative performance in respective graph matters.

[MaxDZ8]

Very interesting experiment... curious to see CubeHash to be so slow.
However I would suggest to iterate at least 1024 calls to rule out possible I$ effects.
Also, clock() does not do what people really wants (it might be equivalent in this context), main problem is the granularity is implementation dependent. I'm surprised it can now be used to bench instructions such as those nowadays.

In any case, don't use it. Use the performance counter or even better C++11 std::chrono.

Groestl should be faster than SIMD on a CPU with AES-NI, I would think.

Are the AES-NI instructions applicable to the bigger Groestl round? I think it's most likely the compiler didn't emit correct code, it seems likely it wouldn't even try. It might be intrinsics or nothing.

[MaxDZ8]

I agree. However, that's what most people use.

[o00o]

It would be much appreciated if someone can confirm as to which algorithms/miners are supported by my Compute Capability 1.1 GPUs such as 9800GTX+/GTS250.  I've tried mining x11 w/ Cudaminer& ccminer on 340.52, the latest driver w/ support for my GPUs but unfortunately, GPU utilization is ~ 1%, making Scrypt my only viable yet unprofitable option. 

Thanks to anyone in advance for offering their help!

[platinum4]

It would be much appreciated if someone can confirm as to which algorithms/miners are supported by my Compute Capability 1.1 GPUs such as 9800GTX+/GTS250.  I've tried mining x11 w/ Cudaminer& ccminer on 340.52, the latest driver w/ support for my GPUs but unfortunately, GPU utilization is ~ 1%, making Scrypt my only viable yet unprofitable option. 

Thanks to anyone in advance for offering their help!

sgminer is for AMD GPU architectures...

[Atomicat]

It would be much appreciated if someone can confirm as to which algorithms/miners are supported by my Compute Capability 1.1 GPUs such as 9800GTX+/GTS250.  I've tried mining x11 w/ Cudaminer& ccminer on 340.52, the latest driver w/ support for my GPUs but unfortunately, GPU utilization is ~ 1%, making Scrypt my only viable yet unprofitable option. 

Thanks to anyone in advance for offering their help!

You're looking for CudaMiner.  You can find the thread here...

https://bitcointalk.org/index.php?topic=167229.0

[o00o]

It would be much appreciated if someone can confirm as to which algorithms/miners are supported by my Compute Capability 1.1 GPUs such as 9800GTX+/GTS250.  I've tried mining x11 w/ Cudaminer& ccminer on 340.52, the latest driver w/ support for my GPUs but unfortunately, GPU utilization is ~ 1%, making Scrypt my only viable yet unprofitable option. 

Thanks to anyone in advance for offering their help!

sgminer is for AMD GPU architectures...

I am well aware that sgminer is optimized for AMD GPUs but given that Cudaminer& ccminer work only w/ Scrypt which is currently dominated by ASICs, the difficulty is now too high for these GPUs to submit any shares even when connected to low difficulty servers.  Despite the performance hit which OpenCL will introduce compared to CUDA, it seems to remain my best option unless someone can point me in the right direction.

It would be much appreciated if someone can confirm as to which algorithms/miners are supported by my Compute Capability 1.1 GPUs such as 9800GTX+/GTS250.  I've tried mining x11 w/ Cudaminer& ccminer on 340.52, the latest driver w/ support for my GPUs but unfortunately, GPU utilization is ~ 1%, making Scrypt my only viable yet unprofitable option. 

Thanks to anyone in advance for offering their help!

You're looking for CudaMiner.  You can find the thread here...

https://bitcointalk.org/index.php?topic=167229.0

Thanks for your input!

[_javi_]

Could somebody plz share the config settings for quarkcoin with 7850/7950 cards?
thanks a lot!

[MysteryX]

Could somebody plz share the config settings for quarkcoin with 7850/7950 cards?
thanks a lot!

I just use the same settings as x11 etc, but I change gpu-threads to 1, and worksize 256.

I wish there was a better quark kernel, it's still wayyyy slower than nvidia cards.

[smokiepot]

there is a better quark kernel. but the person who write this kernel doest make it public. look to the last sites here and you see it.

[_javi_]

thanks for your answers, but i still cant get decent speeds 

[smokiepot]

3.8 mh/s with my r9 280 (without x ). how much you get ? wich hardware , driver , setting ?

[djm34]

new addition to sgminer (github/djm34/sgminer): yescrypt algo
(binaries here: http://ge.tt/5AAmOfF2/v/0?c)

there is 2 implementation of the algo --kernel yescrypt for amd
                                                  --kernel yescrypt-multi for newer nvidia cards (compute 5.2 ie 900 serie for other cards ccminer should be better)

see example.bat for example on how to use it

[JuanHungLo]

new addition to sgminer (github/djm34/sgminer): yescrypt algo
(binaries here: http://ge.tt/5AAmOfF2/v/0?c)

there is 2 implementation of the algo --kernel yescrypt for amd
                                                  --kernel yescrypt-multi for newer nvidia cards (compute 5.2 ie 900 serie for other cards ccminer should be better)

see example.bat for example on how to use it

Thanks for uploading this.  In order to get it to run I had to put a few more .dll in the folder, specifically:

[pallas]

why was the supposedly optimized quark and qubit sgminer post removed?
was it a trojan or a wolf0's leak?

[mitache365]

someone knows the 290x 8GB X11 hashrate ?

[smokiepot]

why was the supposedly optimized quark and qubit sgminer post removed?
was it a trojan or a wolf0's leak?

i think it was because there was a link to another board in there.

[smolen]

why was the supposedly optimized quark and qubit sgminer post removed?
was it a trojan or a wolf0's leak?

The author still distributes them via PM at original Russian forum

aлгo/cк-ть вoльфa/cк-ть/пoтpeблeниe кoмпa c 1 кapтoй, Bт/yвeличeниe пoтpeблeния нa 1 кapтy, Bт
x11/5900/7400/310/10
x13/4600/5700/310/5
qubit/5300/12500/285/35
quark/3000/12900/315/65

(algo/Wolf's speed/speed/Watt/Watt delta)
EDIT: measured on 280x 1100/1500, wattage is for PC with 1 GPU, measurement is not mine

[pallas]

why was the supposedly optimized quark and qubit sgminer post removed?
was it a trojan or a wolf0's leak?

The author still distributes them via PM at original Russian forum

aлгo/cк-ть вoльфa/cк-ть/пoтpeблeниe кoмпa c 1 кapтoй, Bт/yвeличeниe пoтpeблeния нa 1 кapтy, Bт
x11/5900/7400/310/10
x13/4600/5700/310/5
qubit/5300/12500/285/35
quark/3000/12900/315/65

(algo/Wolf's speed/speed/Watt/Watt delta)

is there a linux build? :-)
kernels have been split and parameters changed so you can't just copy the bins over... :-/