Quote from: LoveCats on December 21, 2013, 07:52:05 PM

Quote from: dga on December 21, 2013, 06:56:26 PM

Quote from: LoveCats on December 21, 2013, 06:38:34 PM

Quote from: trell0z on December 21, 2013, 02:37:13 PM

Quote from: LoveCats on December 21, 2013, 01:21:17 PM

Hi,

I've 4x nVidia Quadro K6000 cards and I am running the cudaMiner in autotune mode(if that's the mode in which it runs if no extra flags are specified).

One thing which I've noticed is that, one of the GPU out of 4 gives hash rate of 485kh/s while other 3 ranges from 280-370kh/s. I am getting a total avg of 1450 kh/s. Is this hash rate good at all? How to make the other 3 work on 485 kh/s as well?

Do anyone have any suggestion with manual config to get better result? UI response is of no use to me. Just suggest me the settings that would give maximum performance.

At last, I would like to thank OP for the cudaMiner. As far as I've read, it seems to be only one which makes nVidia cards any useful.

Regards.

Try adding -D to see which config it picks for the cards, and then applying the one that gives the highest hashrate to all of them. After that you can play around with -H 0/1/2. Otherwise from previous posts, you should try the -l config of (amount of smx units enabled on card)x32, which for k6000 should be -l T15x32, or maybe the kepler kernel with K15x32

Thanks bro. The -l with T15x32 boosted the performance. Now getting about 1600 kh/s and all the GPUs are hashing in near about same rate. nVidia GPU's surely suck for mining. This one is such a high-end GPU and such poor performance. Not sure I'll continue this. The trial run's result will decide if its going any further. Thanks again for your help though.

-l T15x32 is the right start. Try T15x64 and see if it runs, but it will probably either crash or give you completely outlandish results. I've been unable to get that config running on a Tesla, though I may still poke at it a little more.

What you'll find, though, is that the Quadro is a terrible ROI for mining: -T15x32 will only use about 2GB of the DRAM on your card (and using more provides no additional benefits). The ratio of registers and FLOPS to DRAM on the Quadro isn't optimal for mining. Lower-end cards with less memory perform similarly (or better because they often run at a slightly higher clock rate). The 780 Ti, for example, has roughly the same compute horsepower but with only 3GB of DRAM, and it's a pretty speedy beast. Well, at least, in the range of nvidia cards.

I'm a little surprised you're not seeing about 40-50kh/s more on each of your cards, but that may be in the noise. Are you running one instance of CudaMiner across all four? You might try running four independent instances if your drivers handle it.

-Dave

Indeed, I was looking at resources and the RAM is so underutilized. Currently, fans are running at 40% speed. Occasionally rises up but pretty much stable near there. Yeah I am running just a single instance for all four.

So as per your suggestion, I should run 4 instance for one card each or 4 instance running on 4 cards each? I believe the later will fry the GPU...lol

I actually did get near about same boost. Previously was getting 1400'ish, now getting 1600'ish hashing rates. So ~200 kh/s increased breaking down to 50 kh/s per card. Before running with manual config, one card would do at much faster and others were doing pretty slow. Now, all are working at pretty much same pace.

One instance per card, specify which one it should run on with the -d operator, -d 0, -d 1 for the next cudaminer etc, I guess you'd have to make 4 folders for that too.
Also, do you have double precision disabled? Supposedly that gives a small boost