CCminer(SP-MOD) Modded GPU kernels.

[chrysophylax]

yep, i dont say that always happens... also all the gigabyte crashed (2 black edition and one normal) had the same problem (capacitor 704),

the ones received from RMA have different capacitors.. no more C704

Also, some cards seems to work better after some months of mining like my MSI 960s

agreed ... they are all different in different environments ...

almost ALL of the gigabyte 7970 / 280x oc cards have failed ... whether they were fans - or burnouts - or leaking heatpipes - or whatever ... so the design needed to be changed quite drastically ...

the nvidia maufacture could have issues with certain batches - but i am just stating that the ones i have had ( the majority from the same batch ) have had no real issues ...

but then again - i use the nvidia at STOCK levels - and dont oc any of them ...

what i also found ( as i used one a while ago ) was that the evga brand of cards were VERY reliable ... both vga cards AND motherboards ...

#crysx

[Slava_K]

On home rig i have Asus cards. Rigs on terminal server pc based on Palit. No fails 2 years...

[chrysophylax]

On home rig i have Asus cards. Rigs on terminal server pc based on Palit. No fails 2 years...

i used to have asus and before that msi ...

they were very good back then ... i moved to gigabyte because they became very reliable cards ...

is there a list that has these cards compared to one another? ...

i would be very interested to see how they compare now ...

not just speeds and oc - but also reliability ...

if it exists at all ...

#crysx

[scryptr]

UPGRADE TO LUBUNTU 15.04--

In the USA it was a holiday weekend for Columbus Day (3-day weekend).  I took advantage of the holiday to upgrade my system to Lubuntu 15.04.  It loaded fairly easily, GTX 970 cards were acceptable to the default Nouveau graphics driver.  I had to specify "nomodeset" prior to launching the install.

First, I went to CUDA Toolkit 7.5.  It was a big mistake!  Neoscrypt ran at less than 1/2 the normal hash rate.  I re-installed Lubuntu 15.04 from scratch, did some web research, and installed the *.run files for first the driver, then the toolkit.  I got an absolutely new beta driver for Linux, v. 358.09, released Oct.12.  I then followed wih the CTK compatible with Maxwell v. 5.2(6.5.19).  I used the *.run files this time because the *.deb files and adding the xorg/edgers PPA may be easier, but the *.run files allow for precise driver selection.  The CTK *.run file, if not run in "silent" mode, asks before installing a v. 343.x driver, the *.deb file does not.

So, I installed the essentials, and cloned ccminer from sp_'s git. I then ran "build.sh" and it bombed.  Apparently, in Lubuntu 15.04, gcc and g++ are both version 4.9, and version 4.8 is required by CTK v. 6.5.   What to do?  I searched the web, and found similar problems occured when CTK moved to v. 5.0.  This is what I did:

I had to install the older versions of gcc, g++.

    sudo apt-get install gcc-4.8
    sudo apt-get install g++-4.8

Check that gcc-4.8 is in /usr/bin/, and same for g++.  Then I could use this solution:

    sudo ln -s /usr/bin/gcc-4.8 /usr/local/cuda/bin/gcc
    sudo ln -s /usr/bin/g++-4.8 /usr/local/cuda/bin/g++

The CTK run file places a symbolic link in the "/usr/local/cuda" directory.  So, the above links direct the compilation to the older v. 4.8 installations of gcc and g++.  Be sure to check that these code systems were installed separately from the current v. 4.9 systems.  Another Linux version may have different directories, but a dedicated miner will dig it.       --scryptr

[bensam1231]

I use 4 x 970 (GV-N970WF3OC-4GD) cards plus 2 x 750 Ti per rig with a single high quality PSU (I refuse to use low quality/low capacity PSUs anymore). The 750 Ti's being anything but Gigabyte OC cards (GV-N75TOC-2GI ) because I had 3 out of 6 of them die on me in about 15 months while having zero issues with Asus and Msi cards. Even though the Gigabyte cards have 36 months of warranty, it take quite some time to replace them.

I've been used Gigabyte GV-N75TOC-2GI since it hit the market. I got 24 of them on 4 rigs. I run them overclock 24/7. Then, 2 of them died (broken fan) during this summer and I noticed that those cards were above the processor. There is extremely hot around that area. If I build a new rig in future, I'll move those cards away from the processor area.

I know a lot of miners like to overclock their cards maximum and they always push the fan faster to lower temp. However, they don't know the fan cannot handle it. In my exp, run at stock clock in the summer and do whatever you like in the winter. For me, I better buy 1 more GTX 750ti than overclock 6 x GTX 750ti to get a extra free card.

That's was not the issue in my case. I had them at 60°C temp target with about 70% fan speed and I don't think they ever reached 60°C and all 3 of them died the same way; computer randomly frooze, after restart the cards weren't recognized at all by the PC and the fans of the cards weren't started either. Also, there was a distinct discoloration (big brown spot) at the back of the PCB on all of them (even on the working ones) at the same spot. It's possible I might just had a bad batch but there's others out there with the same issue which makes me believe these cards are just simply terribly designed with subpar components. Also, I've never had any issues with any other Gigabyte cards (660, 780 Ti, 970) and any other 750 Ti cards (MSI, Asus). Edit: Also, the hottest parts of the cards were only around 70°C, measured with an infrared gun.

Cooling issue. When it discolors, that means there is intense heat above it. Guessing some part of the card wasn't cooled properly (by design or faulty, by design).

Some manufacturers coolers are pretty terrible to be perfectly honest. The only reason they work well is because games don't fully utilize the GPU (usually). You can usually tell who has the best cooler based off throttling that happens. I have EVGA, Gigabyte, and Asus. EVGA definitely seems to have the worse coolers out of the bunch and Gigabyte the best (or are bios flashed in such a way they never throttle). This changes amongst the same generation though. EVGA has like a half a dozen 970s with different versions of coolers and clock speeds so it's impossibly hard to tell what is what. Asus makes one card and Gigabyte two. Windforce I would stay away from as they suck more voltage then the normal model (the ones with the blue LEDs on the side).

I've wanted to try MSI cards, but they always demand a price premium and I'd rather stick to Asus if I'm going to do that.


And yup to the fans dying early if you run them too fast. I ran all my AMD miners with fans at about 80% to maintain cooler GPU temperatures. They always died. XFX in particular had really shitty fans and I had to replace each one of them about twice for the year I owned them. Asus had the best coolers that never went bad and the stock blowers AMD produces also never went bad no matter what speed you ran them at. Funny enough, the cheapest cards I bought from Visiontek ran and overclocked just as well as the other cards and the coolers never failed.

Currently I just run the cards on auto and let them do what they want. If they explode, I RMA them. I have enough room between cards where none of the temperatures look outlandish or start to throttle on their own.

Hey,

I'd like to know what interest there is to buy brand new GPUs (also HDDSs and GPUs) in the community. Through a really big procurement that I'm directing IRL, I'd be able to offer decent discounts. The negative side is that I can only handle a few different models (those my employer chose to buy), although many of each one.

I take two HDDs as examples of what percentage the discount would offer on an item.

WD Green WD60EZRX 6TB IntelliPower 64MB Cache SATA 6.0Gb/s 3.5" Internal Hard Drive Standard price is $279 but newegg have them on discount for $222.

Seagate Desktop SSHD ST4000DX001 4TB 64MB Cache SATA 6.0Gb/s 3.5" Solid State Hybrid Drive Bare Drive Standard price is $279 but newegg have them on discount for $229.

I'd be able to sell these for approx. $170 (EU) and $185 (US) incl. shipping.

More importantly for you, I will be able to offer GPUs (nVidia 970-980ti & the newest AMDs) with the similar discount.


If you are interested, please go here for more info: https://bitcointalk.org/index.php?topic=1207919


Thanks!

None of these are good prices and I'm guessing you're making a healthy profit. Back when I mined Burst, I bought 5TB external drives for $130 a pop. The can now be found for about $115.

[sp_]

Interesting.

I started this thread 17-oct-2014 and now it has got over 400 000 views.  (More than 1000 views per day)

In 3 days we have a 1 year anniversary. I should put a birthday present into release 71.

71 releases in 365 days. A new binary every 1.3 Week.

[pallas]

It looks like cuda 7.5 is ok for x11, quark and lyra, but it's a no no for neoscrypt.
I've tried optimising the cuda code for it specifically, but couldn't get anything better than -20% compared to cuda 6.5.

[sp_]

It looks like cuda 7.5 is ok for x11, quark and lyra, but it's a no no for neoscrypt.
I've tried optimising the cuda code for it specifically, but couldn't get anything better than -20% compared to cuda 6.5.

I managed to improve it to -10% compared to 7.5. (x11) Quark is much worse.

I removed all the maxregcount and all the launchbounds configs and let cuda 7.5 figure it out.

[pallas]

It looks like cuda 7.5 is ok for x11, quark and lyra, but it's a no no for neoscrypt.
I've tried optimising the cuda code for it specifically, but couldn't get anything better than -20% compared to cuda 6.5.

I managed to improve it to -10% compared to 7.5. (x11) Quark is much worse.

I removed all the maxregcount and all the launchbounds configs and let cuda 7.5 figure it out.

Crysx said he installed cuda 7.5 and x11 was a tiny bit faster than with 6.5.
I guess it may be because he's on linux...

[sp_]

It looks like cuda 7.5 is ok for x11, quark and lyra, but it's a no no for neoscrypt.
I've tried optimising the cuda code for it specifically, but couldn't get anything better than -20% compared to cuda 6.5.

I managed to improve it to -10% compared to 7.5. (x11) Quark is much worse.
I removed all the maxregcount and all the launchbounds configs and let cuda 7.5 figure it out.

Crysx said he installed cuda 7.5 and x11 was a tiny bit faster than with 6.5.
I guess it may be because he's on linux...

I have a open supportticket at NVIDIA. The x86 compiler was broken, but they might have improved it in the latest versions.

[Grim]

Pascal has been said to support FP 16 instructions (besides FP32 and FP64).

Is that in any way useful?

[pallas]

Pascal has been said to support FP 16 instructions (besides FP32 and FP64).

Is that in any way useful?

I don't think so. Hashing functions must be precise and always return the exact same value, so they use integer instructions.

[theotherme]

Pascal has been said to support FP 16 instructions (besides FP32 and FP64).

Is that in any way useful?

not really... crypto use mainly integer types...

[chrysophylax]

Interesting.

I started this thread 17-oct-2014 and now it has got over 400 000 views.  (More than 1000 views per day)

In 3 days we have a 1 year anniversary. I should put a birthday present into release 71.

71 releases in 365 days. A new binary every 1.3 Week.

yay ...

that should be a huge milestone ...

lets celebrate it somehow ...

#crysx

[chrysophylax]

It looks like cuda 7.5 is ok for x11, quark and lyra, but it's a no no for neoscrypt.
I've tried optimising the cuda code for it specifically, but couldn't get anything better than -20% compared to cuda 6.5.

I managed to improve it to -10% compared to 7.5. (x11) Quark is much worse.

I removed all the maxregcount and all the launchbounds configs and let cuda 7.5 figure it out.

Crysx said he installed cuda 7.5 and x11 was a tiny bit faster than with 6.5.
I guess it may be because he's on linux...

yup - thats right ...

i did get slightly better hashrates in f22x64c7.5 than with f20x64c6.5 with x11 and quark an lyra2v2 ...

but the latest on git that was just compiled - have gone back to the broken hashrates of f22x64c7.0.28 in x11 - but faster in quark and lyra2v2 ...

i havent tried neoscrypt ... though the x11 is disappointing in speed with the latest commit ...

btw - the stats issue i had with two of the machines is now no longer an issue on those machines since upgrading to f22x64c7.5 ... all the stats ( temp and fan and clocks ) all work ...

#crysx

[sp_]

Looks like the cryptomining blog has made another binary:


http://cryptomining-blog.com/6116-updated-windows-binary-of-the-ccminer-1-5-70-git-fork-by-sp-for-maxwell-2/

This is just the latest from yesterday. Buildt from github.

Small boost in lyra2v2 and quark.

[ldp5500]

Looks like the cryptomining blog has made another binary:


http://cryptomining-blog.com/6116-updated-windows-binary-of-the-ccminer-1-5-70-git-fork-by-sp-for-maxwell-2/

This is just the latest from yesterday. Buildt from github.

Small boost in lyra2v2 and quark.

Slightly slower.

[sp_]

Did you compare with release 70?

Wich card / algo?

[ldp5500]

Did you compare with release 70?

Wich card / algo?

release 70 lyra 2v2 and quark slightly faster.

[sp_]

windows? linux?

750?
750ti?
960?
950?
970?
980?
980ti?