Hi,

With the arrival of Maxwell, we are now supporting 4 generations of hardware, and 6 separate builds of our application on 2 platforms.

This is starting to tax our resources, so some rationalisation is required. I therefore propose the following:

* Remove support for SM1.3
Geforce 200 series cards contribute rather less than 1% of our capacity and, furthermore, lack hardware features that we want to use in future development work.

* Discontinue CUDA 4.2
We need CUDA 6 to properly support the new Maxwell, so I'll be retiring the old CUDA 4.2 build. If you are still receiving version 4.2 WUs, please consider updating your driver to version 343 or above.

These changes will take effect from January 2015

Matt

Very good. I wish more projects would do that.

MJH:

I am in agreement with your proposal.

In fact, the 343+ drivers have dropped support for many of the older GPUs, such that new driver installs will not install drivers for them. In reaction to reading their announcement, a couple months ago I replaced my aging GTS 240 with a 2nd GTX 660 Ti, so that I could continue to install the latest drivers in my milti-generational-GPU rig.

http://nvidia.custhelp.com/app/answers/detail/a_id/3473

This is good news indeed.
I think that crunching on pre-Kepler cards (GTX-2xx, GTX-4xx, GTX-5xx) is a waste of electricity.

RZ,

In the stats I posted a wee while ago, you'll see Fermi era card still constitute 10-20% of our capacity. The lower end 2.1 cards aren't much good for anything other than acemdshort, however.

http://www.gpugrid.net/forum_thread.php?id=3864

Matt

I think that crunching on pre-Kepler cards (GTX-2xx, GTX-4xx, GTX-5xx) is a waste of electricity.

RZ,

In the stats I posted a wee while ago, you'll see Fermi era card still constitute 10-20% of our capacity. The lower end 2.1 cards aren't much good for anything other than acemdshort, however.

http://www.gpugrid.net/forum_thread.php?id=3864

Matt

Matt,

I'm sorry for not being clear, I didn't meant that you should drop support of the Fermi based cards.
I've intended my message to encourage my fellow crunchers to upgrade their GPUs, as the Fermi based cards became inefficient, running them in the long term is *not* worth its costs anymore.

Zoltan

Hi, Matt:

I understand the reason for this rationalization. At present I process GPUGrid tasks with two NVIDIA GeForce GTX 650 Ti (1024MB) GPUs and driver 334.89. I have no intention of spending any more on my PCs: can you please let me know when you will discontinue providing work for this GPU.

Thanks,

John

Hi,

With the arrival of Maxwell, we are now supporting 4 generations of hardware, and 6 separate builds of our application on 2 platforms.

This is starting to tax our resources, so some rationalisation is required. I therefore propose the following:

* Remove support for SM1.3
Geforce 200 series cards contribute rather less than 1% of our capacity and, furthermore, lack hardware features that we want to use in future development work.

* Discontinue CUDA 4.2
We need CUDA 6 to properly support the new Maxwell, so I'll be retiring the old CUDA 4.2 build. If you are still receiving version 4.2 WUs, please consider updating your driver 343 or above.

These changes will take effect from January 2015

Matt


[/quote]

John,

At present I process GPUGrid tasks with two NVIDIA GeForce GTX 650 Ti (1024MB) GPUs and driver 334.89.

Those cards are just fine, and will be useful for a long time to come. It's only he Geforce 200-series we need to wave good-bye to.

If you could update your driver to 343 sometime before the new year, that'd be appreciated. You'll get the newer CUDA 65 app, which will - eventually - be faster than 60.

Matt

I'm crunching on an NV Optimus notebook. The latest repository drivers running Gpugrid are 331.38 and does Cuda4.2. Official drivers have problems with Optimus. Seems time to switch to CPU projects is coming. Rosetta probably...

I've had a GTX470 sitting on the shelf for some time. It's still a good GPU, but for crunching it just uses too much power relative to the 600 and 700 ranges, never mind the 980 and 970.

Going by the official GFLOPS/Watt (Single Precision) you can see a big jump from Fermi to Kepler:

GTX 	GF/W (by series number)
980 	28.0
780 	15.9
680 	15.85
580 	6.48
480 	5.38

GTX 	GF/W (by most efficient in series)
980 	28.0
780Ti 	20.2
690 	18.74
560Ti 	7.42
480 	5.38

Didn’t include the 750Ti (21.8) because it’s a Maxwell.

In terms of running costs, the GTX980 is 3.77 times more efficient than the GTX560Ti (the most efficient 900 series GPU vs the most efficient 500 series GPU). The 980 is even 40% more efficient than a GTX780Ti.

For those with a budget the running cost of a GTX970 (24.1) is 3.24 times more efficient than a GTX560Ti, and the 750Ti (21.8) is 2.93 times more efficient than a GTX560Ti.

So it would be in the interest of most crunchers to stop using anything below the GTX600 range, and as Zoltan pointed out, worth encouraging. I suggest giving notice to drop support for 500's in ~6 months (March 2015).

---

FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

I agree with Skgiven and RZ.
Once GM204 Maxwell's driver support fully integrates filtering technologies support- Molecular programs like PyMOL/JyMOL will be able to view fragments(atoms) and residues (amino acids) like never before. Maxwell's are certainly helping/advancing science in many ways.

On a side note: mobile Maxwell were released, and have very curious core counts. A future GM206/desktopGTX960? GTX970m=1280c/10SMM/3GPC/80TMU/48ROP
(Future GTX960?) A GTX980m(1536core/96TMU/64ROPS) is 12SMM/3GPC rather 13/4GPC for desktop GTX970.
This is pure speculation on my part: I think NVidia is finding Maxwell GPC (4SMM/32TMU/16ROPS) is unable to disable SMM in similar fashion to SMX, due to the 32 core subsets with a separate warp. Unlike SMX monolith design where SMX shares all cores, SFU, LD/ST units with warp. Kelper SMX has (8 separate dispatch-->one Large crossbar-->separate [4] issue for a subset(32c) and [1] issue for 32SFU and [1]issue for 32LD/ST.) Maxwell breaks up dispatch into four pairs of two feeding a separate crossbar for separate issue to subset of 32c and issue for 8SFU,8LD/ST. (SMM= 4 issues for SFU/ 4 issues for LD/ST/ 4 issues for [4] 32 core subset.

GTX GF/W (by most efficient in series)
980 28.0
780Ti 20.2
690 18.74
560Ti 7.42
480 5.38
[/pre][/list]
Didn’t include the 750Ti (21.8) because it’s a Maxwell.

In terms of running costs, the GTX980 is 3.77 times more efficient than the GTX560Ti (the most efficient 900 series GPU vs the most efficient 500 series GPU). The 980 is even 40% more efficient than a GTX780Ti.

For those with a budget the running cost of a GTX970 (24.1) is 3.24 times more efficient than a GTX560Ti, and the 750Ti (21.8) is 2.93 times more efficient than a GTX560Ti.

Has anyone actually measured the power draw from various 970 and 980 cards while running GPUGrid? We know the 750Ti is running at about 60 watts. According to the tom's Hardware article linked below 970/980 power consumption while running CUDA tasks may be rather high. Seems we need to see some actual power draw figures for various cards while running GPUGRID tasks.

Tom's Hardware power consumption dissertation for the 970/980:

http://www.tomshardware.com/reviews/nvidia-geforce-gtx-980-970-maxwell,3941-12.html

For those with a budget the running cost of a GTX970 (24.1) is 3.24 times more efficient than a GTX560Ti, and the 750Ti (21.8) is 2.93 times more efficient than a GTX560Ti.

So it would be in the interest of most crunchers to stop using anything below the GTX600 range, and as Zoltan pointed out, worth encouraging. I suggest giving notice to drop support for 500's in ~6 months (March 2015).

That is a very nice comparison, and I am sure quite useful for those who have not done the tests themselves. (It is no accident however that I now have six GTX 750 Ti's, which I picked up heading into the summer.)

However, my concern is not trying to tell people how to spend their money or force them into more efficient cards, useful though that would be for some purposes. (The operating costs are different in the U.S. than in Europe, for example.) My concern is compatibility and the effort that MJH and others must expend to keep the older cards operating. I have seen many projects where problems arise trying to keep a few crunchers happy with their present hardware, even though that detracts from the science of the overall effort. I think we should keep focused on why we are here.

GTX GF/W (by most efficient in series)
980 28.0
780Ti 20.2
690 18.74
560Ti 7.42
480 5.38
[/pre][/list]
Didn’t include the 750Ti (21.8) because it’s a Maxwell.

In terms of running costs, the GTX980 is 3.77 times more efficient than the GTX560Ti (the most efficient 900 series GPU vs the most efficient 500 series GPU). The 980 is even 40% more efficient than a GTX780Ti.

For those with a budget the running cost of a GTX970 (24.1) is 3.24 times more efficient than a GTX560Ti, and the 750Ti (21.8) is 2.93 times more efficient than a GTX560Ti.

Has anyone actually measured the power draw from various 970 and 980 cards while running GPUGrid? We know the 750Ti is running at about 60 watts. According to the tom's Hardware article linked below 970/980 power consumption while running CUDA tasks may be rather high. Seems we need to see some actual power draw figures for various cards while running GPUGRID tasks.

Tom's Hardware power consumption dissertation for the 970/980:

http://www.tomshardware.com/reviews/nvidia-geforce-gtx-980-970-maxwell,3941-12.html

I looked at power consumption of my 980 on F@H. Below are my data and also a link to the thread. I will do the same sort of test on GPUGrid this weekend.

"I have a Kill-a-Watt power meter at the wall on this 64-bit linux mint 17 LTS system:

Intel 3930K, Asus x79 sabertooth MB, Gigabyte GTX980 Nvidia reference design (Model #: GV-N980D5-4GD-B)

Power readings (SMP-10, project 6348, GPU core 17, project 9201)

Full load with 980@1252MHz (no overclock): TPF (1 min 38 sec), 387 watts total power
Full load with 980@1452MHz (+200 MHz overclock)): TPF (1 min 28 sec), 401 watts total power
GPU Folding on Pause (980@135 MHz): 263 watts total power
Both GPU and SMP folding on pause: 114 watts total power

980 from Idle to 1252MHz (no overclock): 124 watts
980 from Idle to 1452MHz (+200 MHz overclock): 138 watts

I'm not sure how much power the 980 is drawing at idle, but these numbers look pretty good to me."

https://foldingforum.org/viewtopic.php?f=38&t=26757&start=15

Bumping, since first announcement had no effect on reducing 4.2 app contribution.
Possibly crunchers who never read the forums..

Bumping, since first announcement had no effect on reducing 4.2 app contribution.
Possibly crunchers who never read the forums..

Hi Matt, you can use the "Notices" in BOINC Manager to do the announcement and perhaps repeat that every fortnight.

---

Greetings from TJ

Hi Matt, you can use the "Notices" in BOINC Manager to do the announcement and perhaps repeat that every fortnight.

Neat, thanks. Can you see the notice now?

Matt

Hi Matt, you can use the "Notices" in BOINC Manager to do the announcement and perhaps repeat that every fortnight.

Neat, thanks. Can you see the notice now?

Matt

Yes I can see it in notices. Keep up the great work

I've got a GeForce 760 card, and the driver is 344.48. I know "every bit counts," but it makes sense that you're moving away from older hardware. Eventually the older hardware just becomes a waste because newer hardware can do it faster, better, and more effecient.

---

Hello.

What a lot of you don't realize is everyday there are fewer and fewer people using BOINC. Maybe users are stable on GPUGRID, however, overall people see to get bored, even long time users and eventually stop running BOINC.

So, the want of some users to cut off anymore than what the admins have suggested based off efficiency is just wrong. Sure it'd be ideal if everyone had the most efficient cards. But right now, BOINC is sort of spasiming due to lack of new users to make up for the ones leaving. They need a good PR campaign or something. Or some kind of better reward system (maybe actually giving something like bitcoins for use, such as GRIDcoins). Anyway - cutting off users due to inefficiency is a bad idea. It's alienate the people and would not be worth making the world a slightly better place. I don't think I'd return to the project if it did that to me, seeing how many other ones there are out there that wouldn't do that to me.

Everyone who reads this ought to try to get a friend whose never used BOINC to try it, rather than worry about inefficient GPUs on the project. For real.

Regards,

Matthew Van Grinsven