Topic: [ANN] cudaMiner & ccMiner CUDA based mining applications [Windows/Linux/MacOSX] - page 1127. (Read 3426947 times)

Another quadro (4000) was getting around 24 on the previous drive, now at 48 configured at 68x2. When I ran without a specifed kernal, it configured at 68x7 but was hashing junk shares. so I started at 68x1, which did a about 14 kh/s, then tried 68x2 and it got up to 48 kh/s. Anything hire produces junk shares or kills cudaminer and crashes the driver. I wish I knew more about this kernel setting... would play around with it a bit to see what would work. Instead I just crash my system if I mess with it too much :| I would be alright with that if I was physically near the machine to restart it.

Quadro 4600 was getting around 20kh/s in 04/09's version, averaging closer to 24 with spikes up to 25 and 26 now. I have a machine at the office with a 600 in it as well.. don't have access to it right now, but I'll check it out in the AM and provide the stats. The 4600 is running at 47x2, I think it was going at 24x4 before.

18khash/s from a overcloked GT520 920/810.

For this reason I've put up the 04-09 version again in the first post. Should reach 290 kHash when slightly overclocked.

What's with GPU 2,3,4,5 in the above screenshot, all showing garbage names? Trying to run 6 threads here? Use 2

About the loss of GPU perf when maxing out the CPU: the CPU is doing preparatory work for the GPU by doing SHA 256 hashes. If you slow down this CPU operation, the GPU runs out of work to do.

Future cudaminer version will at least try to minimize the load on the CPU by re-introducing SSE2 vectorization. For the moment I've left it out because it simplifies the compilation on Windows when not having to deal with assembly code.

Christian

doesn't work on my titans either atm.  look forward to new versions. 

since you have an i7, why not switch to the internal Intel HD graphics for your primary display and then let the nvidia cards run their thing in the background.  This will eliminate any drops you see from moving the mouse, scrolling, moving windows, etc.  

...unless of course it's not a dedicated mining machine and you use it for games or other 3d work.  As for your loaded cpu dropping your gpu hashrate, I have no ideas as we have the same cpu and I don't experience that.


I've got the following and haven't experienced anything that you've mentioned (including the effects of a fully loaded cpu dropping your hashrate). - maybe it's because I'm using the intel hd graphics as my primary display.
i7-3770k @ 4.6Ghz
Z77 chipset
24GB ddr3 1800
2x GTX560SE
1 GT430
Windows 7 (Aero disabled)
driver 314.22

I get a total of ~240Kh/s from the gpu's.  43Kh/s from the CPU (6 threads).

About the currently broken Titan support:  I am beginning to think that it is a nVidia compiler bug. I found a posting on the nVidia forums stating where someone was complaining that his Release builds would crash when targeting the compute 3.5 architecture, whereas the same code runs just fine when built for previous architectures. His code showed no issues when run through cuda-Memcheck.

I just took my Titan kernel code (except for the hardware specific funnel shifter part), plugged it into compute 1.0 architecture and ran it on my laptop's GPU with no problems. It appears that I am having the same issues as the guy I just mentioned.

Christian

Here we go!! Finally I can post!!

i7 3770
H67
660Ti
16GB ram
Windows 7
Driver 314.22

no OS optimization (even Steam is running in background)
bad-bad-bad ventilation (Cooler Master Elite mini-itx)

with Stratum_Proxy feeding (everything seems smoother, highly recommended):

2013/04/10: 159.45 khash/s - conf 42x7
2013/04/09: 132.50 Khash/s - conf 147x2

Some hints I like to verify with you:
- I managed to get 24 khash extra using the CPU, forget the Hyperthreading and setup 3 threads (if you got 4 cores), I've no hints for dual cores. Sticking to 4 or 8 threads (full cpu load) dropped the GPU at about 100 Khash. Giving 7 threads (so using HT) had basically the same bad impact
- don't mess with the console window (drag, tab, scroll, etc), just scrolling dropped the rate of a chunk from 150 to 110. After the test (keep the mouse steady!) stick to your pool dashboard to see the results (in my case 147 over 30 jobs - test done today)

I'd like to test:
- vanilla desktop (no Aero interface)
- different set of drivers (in this case I should setup a dedicated partition and format everytime - I'm too lazy, I must admit it)
- Linux

...meanwhile I'm relocating to a new home, so I' really don't know when I'll be able to do that :-|

Two last words:
- it's an awesome work, I've mined only 0,2 LTC of which 0,1 was burned by the pool for the transfer to my wallet (damn testing!), otherwise you already had had a donation from me too
- the software is really SOLID. It ran almost 24 hours (from an itx computer!) without any hiccup. Meanwhile Minerd had two crashes!

Windows  based version doesn't work for me. it will not load the cuda module, and no one seems to have the answer. I've googled it and checked the forums.

There are a gajillion other options for SHA based mining, including a SHA miner for CUDA (https://bitcointalksearch.org/topic/nvidia-kepler-k20-from-134mhashs-to-330mhashs-with-cuda-163750). There's no reason for the author to even give any time to this

Terracoin used SHA, and others. Which really don't have a lot of power behind them.

This is a RTFM situation. The manual states SHA hashing is done one the CPU. Why would I optimize for Bitcoin mining, if you would be competing against FPGAs and ASICs?

For CUDAMiner version 10/04, GTX650 Ti on Windows
Litecoin mining at ~80 khash/s, autotuning takes about 3~5 minutes.

There seems to be some problems with Bitcoin mining though (-a sha256d):
1. NULL displayed for GPU device (regardless of device specified)
2. very slow speed (~1 Mhash/s) compared to other current miners (e.g. OpenCL with poclbm, and CUDA with rpcminer-cuda ~80 Mhash/s)

I don't think my gtx260 saw a change either.  My fermi cards did though...a very nice change in fact.  ~30kh/s change.

I concur, this is a step backwards. But on my GTX 260 I do not see the same regression. In fact the card is the same generation as yours. Strange.

I have received a screenshot of a single Geforce titan doing 280 kHash/s with the April-09 release of cudaminer.
Use a dual Titan system and you will be hashing at 560 kHash.   You will have to mine for a looooong time to get your expenses back, though.

Christian

Slightly disappointing, I find. I have a 8800GT (112 shaders) also, maybe I will find some time to plug it in and test it. But now I will be torturing my new arrival, a GT 560Ti.

Christian

8800GTS 320MB:

Startup failed due to missing MSVCR100.dll, downloaded it here.
Restarted succesfully after installation. Autotune output: http://pastebin.com/TGK6JZLm

Settles at 17-19KH/s. This card is normally mining bitcoins at about 18 MH/s.

A few autotune reruns show that it always uses either S35x1, or S36x1. Both resulting in 18-19KH/s. Not bad, for a granny. 

new uptade working greatly even on mid-end mobile devices, jumping from 20Kh/s to 26Kh/s on my gt 540m
autotune won't start though, and --no-autotune sets it to 16x2 (which i don't understand what it means)

now i'm wondering how far optimization will bring us

Just tried the 2013-04-10 version on my GTX570. Upped the speed for me too.
Autotune output: http://pastebin.com/YgtHLWeC. Selected config 30x7 this time, promising around 180KH/s. After a while it is doing 175-178KH/s, pool reports 172.

GTX570 - 30x7
Win 8 x64
314.22 WHQL driver
stock: 732 core, 1900 shader

After 10 minutes, still no stales (yay!!!) 


The 2013-04-09 version gave me config 120x2 most of the runs, and did 145-150KH/s.

I have an old PC laying around here too, with a 8800GTS 320MB.  Will try to get that online too.