The attached shows the latest condition of the two cards. GPU1 is the EVGA, GPU2 is the MSI. Since the last post I have lowered the voltage on GPU1 another 'tick' to 1.063. Besides that, the room temperature is down from 22-23 C during the day, to about 18 C now. (It is currently a balmy 13 F outside, forecast to go to 9 F. This 1920s apartment building is not terribly well insulated. You can feel cold air washing down the walls which are on the periphery.)

Consequently, the cooler temps are partly from cooler air feeding in. However, it is more telling that the upper and lower cards are running at almost the same clocks, and are holding the same temperature. I'm surprised, after its initially disappointing cooling performance, that I've gotten the EVGA/Zalman combination down this far, especially considering the relatively high voltage it is still running at.

So the Zalman turns out to be a fairly decent cooler, if it is properly interfaced to the chip. I still wish that it had automated fan control instead of the cheesy cable running out the back of the case to a little manual controller. I may try inserting a temperature sensor in the hottest part of the heatsink, so I can put it under motherboard control.

Could a kinder mod please restore the title of this thread to "Trial division with CUDA (mfaktc)", and would all mods please stop changing thread titles of threads whose forum is displayed above the Soap Box on the default front page?

[QUOTE=Dubslow;395397]Could a kinder mod please restore the title of this thread to "Trial division with CUDA (mfaktc)", and would all mods please stop changing thread titles of threads whose forum is displayed above the Soap Box on the default front page?[/QUOTE]

Seconding the request. I understand it's part of the forum culture, but important threads should not be renamed without a good reason (such as making the title more descriptive). Otherwise, it could easily cause problems for people who are searching for serious threads. At the very least, the original title should be part of the new one to avoid confusion:

[QUOTE]ChemTrails division with CUDA (mfaktc) [was: Trial division with CUDA (mfaktc)][/QUOTE]

To be honest, the frivolous title changes have stopped being funny a long time ago for me.

Some of them, especially "down south", can be quite funny. :smile: Anyways, I guess the current title is "good enough". It's more accurate but less historical.

[QUOTE=ixfd64;395445]ChemTrails division with CUDA (mfaktc) [was: Trial division with CUDA (mfaktc)].[/QUOTE]
It wasn't.
It was "Trail division with CUDA (mfaktc)" for many months, if not years.

good news, everyone!
 

So a new topic for this old thread? So here is a new version of mfaktc. Evolution, not revolution!

Highlights (for full changes check Changelog.txt):[LIST][*]added support for Wagstaff numbers: (2[SUP]p[/SUP] + 1)/3[*]added support for "worktodo.add"[*]enabled GPU sieving on CC 1.x GPUs[*]dropped lower limit for exponents from 1,000,000 to 100,000[/LIST]
Special thanks goes to Jerry Hallett who spent a lot time on building and testing Windows binaries aswell as some hints for Windows compability. [B]Thank you Jerry![/B]

I guess it is a little bit faster than 0.20 for the current Primenet wavefront on most, if not all, cards. [B]Please post your mileage.[/B]

As usual: finish your current assignment with your current version of mfaktc, mfaktc will refuse checkpoint files from other version than itself... yes I know - I'm paranoid in this case!

So I know what you want now: here are the files:[LIST][*]source code: [url]http://www.mersenneforum.org/mfaktc/mfaktc-0.21/mfaktc-0.21.tar.gz[/url][*]basic Windows package ([I]regular[/I] 32- and 64-bit mfaktc) - CUDA 6.5: [url]http://www.mersenneforum.org/mfaktc/mfaktc-0.21/mfaktc-0.21.win.cuda65.zip[/url][*]for those who (think) they need them - additional binaries for Windows - CUDA 6.5: [url]http://www.mersenneforum.org/mfaktc/mfaktc-0.21/mfaktc-0.21.win.cuda65.extra-versions.zip[/url]
Download the basic package and than ADD these binaries to the folder. Includes [I]LessClasses[/I] builts and binaries for [I]Wagstaff[/I] numbers[*]Linux binaries - CUDA 6.5: [url]http://www.mersenneforum.org/mfaktc/mfaktc-0.21/mfaktc-0.21.linux64.cuda65.tar.gz[/url]
compiled on openSUSE 12.2[/LIST]
What's next (read: mfaktc 0.22)?[LIST][*]remove support for CC 1.x GPUs, Nvidia drops support for CC 1.x in CUDA 7.0 so I'll do so, too! Will remove lots of [CODE]#if (__CUDA_ARCH__ >= FERMI)[/CODE] in the code.[*]remove 71bit kernel, only choosen for CC 1.x GPUs, less code to manage![*]remove support for older CUDA versions, will remove lots of [CODE]#if (CUDART_VERSION >= ????)[/CODE] in the code. I'm not sure which version I'll require as baseline, perhaps I go as high as CUDA 6.0 or even 6.5... unsure here.[/LIST]
So 0.22 will help coders, no users (directly)... sorry!

Oliver

Fantastic! Great work.

[QUOTE=TheJudger;395689]for those who (think) they need them - additional binaries for Windows - CUDA 6.5... Includes [I]LessClasses[/I] builts[/QUOTE]Thank you! :max::bow::w00t::banana:
I can now run 1 instance per GPU instead of 4 per GPU (on >1000M TF to 64)

Actually on my GTX 580 I'm running 2 instances for now since each assignment takes 0.5s, but in truth the throughput isn't much different despite filling my GPU to 100% instead of 70%. I'll probably go back to 1/GPU when I plug a new 570 into that system later this week.

[QUOTE=TheJudger;395689]remove support for CC 1.x GPUs
remove support for older CUDA versions[/QUOTE]I'm totally fine with that. I'd be perfectly satisfied if you called 0.21 the last release for old CUDA and GPUs, and made 0.22+ for CUDA 7.0 only. Would probably greatly simplify your code and it's not like those who are stuck on old GPU and/or CUDA can't use mfaktc anymore, just no higher than 0.21

Does this mean we can expect mfaktc 1.0 in the not-so-distant future?

For anyone compiling on Ubuntu 14.04 with the packaged CUDA 5.5 provided by Canonical, I had to make these changes:

CUDA_DIR = /usr/
CUDA_LIB = -L$(CUDA_DIR)/lib/x86_64-linux-gnu/

And because I don't have a Maxwell, and this was failing with CUDA 5.5:

#NVCCFLAGS += --generate-code arch=compute_50,code=sm_50 # CC 5.x GPUs will use this code


Anyway, the new version sped up my cards by 1%:

GTX 580: 424 -> 430
GTX 580: 428 -> 433
GTX 760: 259 -> 261

on DCTF 41M 70->71 work. I haven't checked the difference on the GT 430's and GT 520's.

Excellent!

[QUOTE=ixfd64;395699]Does this mean we can expect mfaktc 1.0 in the not-so-distant future?[/QUOTE]

*hmmmm* [B]NO![/B]
What's so special in version 1.0?

Oliver

P.S. check README.txt, chapter "7 FAQ" for version numbers :smile: