Huge different in hashrate for v1.03 and >v2

[kutysam]

I'm using an old card. 660ti. Compute 3.0 capability Hashrate differs greatly when I do a benchmark with the same launch config.

Hashrate on >v2.0 --> ~68H/s

C:\New folder>ccminer --version
    *** ccminer-cryptonight 2.00 (64 bit) for nVidia GPUs by tsiv and KlausT
    *** Built with Visual Studio 2015 using the Nvidia CUDA Toolkit 8.0

 tsiv's BTC donation address:   1JHDKp59t1RhHFXsTw2UQpR3F9BBz3R3cs
 KlausT's BTC donation address: 1QHH2dibyYL5iyMDk3UN4PVvFVtrWD8QKp
 for more donation addresses please read the README.txt
-----------------------------------------------------------------
ccminer-cryptonight 2.00
libcurl/7.50.3 WinSSL WinIDN

C:\New folder>ccminer.exe --benchmark -l 4x84
    *** ccminer-cryptonight 2.00 (64 bit) for nVidia GPUs by tsiv and KlausT
    *** Built with Visual Studio 2015 using the Nvidia CUDA Toolkit 8.0

 tsiv's BTC donation address:   1JHDKp59t1RhHFXsTw2UQpR3F9BBz3R3cs
 KlausT's BTC donation address: 1QHH2dibyYL5iyMDk3UN4PVvFVtrWD8QKp
 for more donation addresses please read the README.txt
-----------------------------------------------------------------
[2017-10-26 10:44:42] Using JSON-RPC 2.0
[2017-10-26 10:44:42] 1 miner threads started, using 'cryptonight' algorithm.
[2017-10-26 10:44:42] GPU #0: GeForce GTX 660 Ti (7 SMX), using 84 blocks of 4 threads
[2017-10-26 10:44:52] GPU #0: GeForce GTX 660 Ti, 72.27 H/s
[2017-10-26 10:44:52] Total: 72.27 H/s
[2017-10-26 10:45:02] GPU #0: GeForce GTX 660 Ti, 65.91 H/s
[2017-10-26 10:45:02] Total: 65.91 H/s
[2017-10-26 10:45:07] GPU #0: GeForce GTX 660 Ti, 65.95 H/s
[2017-10-26 10:45:07] Total: 65.95 H/s
^C

Hashrate on v1.03 ~ 250h/s

C:\New folder>ccminer-cryptonight.exe --version
    *** ccminer-cryptonight 1.03 (64 bit) for nVidia GPUs by tsiv and KlausT
    *** Built with Visual Studio 2013 using the Nvidia CUDA Toolkit 8.0

 tsiv's BTC donation address:   1JHDKp59t1RhHFXsTw2UQpR3F9BBz3R3cs
 KlausT's BTC donation address: 1QHH2dibyYL5iyMDk3UN4PVvFVtrWD8QKp
 for more donation addresses please read the README.txt
-----------------------------------------------------------------
ccminer-cryptonight 1.03
libcurl/7.50.1-DEV WinSSL WinIDN

C:\New folder>ccminer-cryptonight.exe --benchmark -l 4x84
    *** ccminer-cryptonight 1.03 (64 bit) for nVidia GPUs by tsiv and KlausT
    *** Built with Visual Studio 2013 using the Nvidia CUDA Toolkit 8.0

 tsiv's BTC donation address:   1JHDKp59t1RhHFXsTw2UQpR3F9BBz3R3cs
 KlausT's BTC donation address: 1QHH2dibyYL5iyMDk3UN4PVvFVtrWD8QKp
 for more donation addresses please read the README.txt
-----------------------------------------------------------------
[2017-10-26 10:45:22] Using JSON-RPC 2.0
[2017-10-26 10:45:22] 1 miner threads started, using 'cryptonight' algorithm.
[2017-10-26 10:45:22] GPU #0: GeForce GTX 660 Ti (7 SMX), using 84 blocks of 4 threads
[2017-10-26 10:45:25] GPU #0: GeForce GTX 660 Ti, 243.28 H/s
[2017-10-26 10:45:25] Total: 243.28 H/s
[2017-10-26 10:45:30] GPU #0: GeForce GTX 660 Ti, 254.43 H/s
[2017-10-26 10:45:30] Total: 254.43 H/s
^C

Hope you can take a look at this. Thanks!

[kutysam]

oh, i'm using Cuda 9.0 (latest driver from nvidia site 388)

[KlausT]

I have no idea why version 2.00 is so slow. Would you please try version 2.06beta ?

[kutysam]

I've tried version 2.06, same as 2.00. In addition, I've also tried xmr-stak-nvidia and others. All have poor hashrates with their latest version. At least this shows that the error lies in one of the commits from 1.03 ~ 2.00 https://github.com/KlausT/ccminer-cryptonight/commit/e60463a31ab8c3d0fff90162632f0ca3fcc269c3

My system is a windows 10 64bit, nvidia 388.00 cuda driver 9.1.56, 660ti 2gb Compute capability(version): 3.0

Do let me know if I can assist in any way, perhaps you could post some binaries and I can test them out?

[KlausT]

Could you try this? ccminer-cryptonight-test.zip This is the commit "less structs, a little bit more speed" By the way, I wonder why nobody else has complained yet.

[kutysam]

Still the same issue with this file.

However, there are a few other settings where i can get almost the same hash rate as the 4x84, but with the latest version. Example: --bfactor=10 --bsleep=30 -l 20x28 ---> gives me 250H/s

I guess that's why many didn't complain. But it's just quite shocking to see a huge drop when 4x84 was selected for the latest configs.

Anyways, I guess you can keep this a low priority as its an old card and there are other launch configs which give similar hash rates.

[kutysam]

Anyways, here's a side question, could you increase the limit for bfactor? If i'm not wrong, xmr-stak-nvidia has a higher limit for bfactor.