Hi,
we would like to update the ACEMD application with a newer version which is faster and offers some extra features for the science.

We will try to keep only 4 versions:
compiled for CUDA3.2 for windows and Linux
compiled for CUDA2.2 for windows and Linux

However, we are not sure that the app will compile for CUDA2.2. I will confirm in the next few days.

gdf

Confirmed.
We will start to come out next week with cuda2.2 applications for windows and linux in beta.

If all goes well we will go on with the cuda3.2. All cards will be able to receive cuda3.2 apps.

gdf

For us it's the same speed.

By the way, we will deprecate cuda3.1 and 3.0 applications due to several bugs in the CUDA libs. So you need to change driver to the latest driver to keep receiving cuda3 applications. This will be done slowly and notified in the news.

gdf

GDF, can you confirm that the latest driver will only be needed for Fermi cards?

The v260.89 and v260.99 drivers are bad for the 200Series cards, and probably earlier cards - the frequency drops and stays low.

I have not noticed any drop in frequency for my GTX470's with the latest driver, so I expect 260.99 works for Fermis.

I assume that if you have a g200 cards then you are still using the cuda2.2 application at the moment. In this case, you don't need to change driver.
You will need to change driver to receive a CUDA3 application. A 3.1 driver will not be enough when we remove the 3.1 app.

gdf

Thanks.

I guess this might in itself bring Linux support to the GTX460, GTS450 and still unreleased GTX475 (basically just a GTX460 with a full complement of shaders and cores) as it needs the 260.99 driver, but I don’t know what is under the bonnet of the latest ACEMD tool, so we might still need to wait on the 3.2 dev app to move from RC to final version?

Whatever the case, if it improves reliability or increases speed it will be very welcome.

For CUDA 2.2 you don't need to update your driver.

For CUDA 3.2, who knows, but this only applies to Fermi's and you will need the latest driver - whatever that will be tomorrow?

I don't know why I am surprised NVidia said 265.xx, it's not as if they are conformists when it comes to time tables, design, testing, series... Still it is a bit of a curve ball. If they do manage to ship a GTX475 by the end of Nov, the driver scene will be very messy. Chaos looms if they inflict a GF110 upon us.

What driver level is needed for cuda 3.2 support?

Since 3.2 is not yet released, that is an unanswerable question.

MJH

Found out myself:

nVidia has released the Forceware 261.00 driver. This driver is a special release intended primarily for developers, providing support for the CUDA Toolkit 3.2 RC2, so it's best used for experimental purposes.

It seems the fist OFFICIAL cuda 3.2 drivers will be the 265.xx series...

Here's some info on cuda 3.2:

http://developer.nvidia.com/object/cuda_3_2_toolkit_rc.html

There was a link to the v261.00 drivers but it has been removed.

Developer Drivers for WinXP (261.00) <Coming Soon!>
Developer Drivers for WinVista and Win7 (261.00) <Coming Soon!>
Saw it listed here

I have changed the server policy.
Now all cards with the appropriate driver should be able to get the CUDA3.1 application. G200 cards installed together with Fermi cards will not work properly until we will release a cuda3.2 app.

This allows many old cuda2.2 bugs to be fixed on g200 cards if you just update the driver to a recent one. Change only if you have problems with the cuda2.2 app.

gdf

To test if this frequency bug effects all 200series cards or just some I installed 260.99 for a GT240 (DDR3 1GB version) on a W7x64 system.
Boinc downloaded the CUDA3.1 app and is now running a 6.11 TONI_KKi4 task.

It's possible it will not cause problems on the 40nm GT240s and may only cause problems on specific makes of cards, depending on their Firmware.

GPU details:
Core Clock at 597MHz,
Shaders at 1595MHz,
RAM at 800MHz (x2)
Fan speed 60%
GPU temp 49 deg C.
GPU usage 89 to 94 %, mostly 91%.

No problems so far; the GT240 has not slowed any after 1h.

Picked up a 6.11 CUDA 3010 task on another GT240 with the 258.96 driver, so we dont need to use the 260.99 driver for 200 series cards.

Thanks Beyond, that paints a clearer picture.
It looks like the 260.89 and 260.99 drivers may work for 200series cards on Win7 and Vista, but not on XP. I suggest people only use the 260.99 driver, if they must, but we really need a better driver before going to CUDA 3.2 only, if that is the plan. Hopefully NVidia will bring out a fully functional 261.xx soon.

I will update to the 260.99 driver on my Vista system this evening (UK) and give it a go. It has 4 GT240's.
For some still unknown reason my GTX260 failed all tasks on Vista and W7 the last time I tried, but I may try it on XP X86 in a day or so (if no new drivers turn up). It's sitting on my desk awaiting a test system.

It might, but other updates did not, so I'm not holding out much hope. Give it a go in a few days and see.

Would the updated application solve the problem regarding on the 65nm GTC260s?

I also picked up 6.11 tasks on both of my GTX 260 cards with 258.96 drivers.

The 6.11 app so far is running about the same speed on the GTX 260, but all 3 of my GT 240 cards have slowed down about 60%. GPU usage is still above 90% but the WUs are running so slow that it's not worth keeping them on the project with the 6.11 app. Can we PLEASE go back to 6.05?

Yes, there will be a CUDA2.2 app for Windows. In fact, there is already one as long as you have a cuda2.2 driver.

gdf

I hope GPUGRID will have also in future a CUDA V2.2 app for Windows.
I need to stay at nVIDIA driver 190.38, the fastest driver for/on my systems.


(I tested with app_info.xml file, CUDA V2.3 .dll's the GPUGRID CUDA V2.2 app.
It's look like it was little bit slower then (but wasn't a real bench-WU test). But maybe because of the app wasn't compiled with CUDA V2.3 ..)


But I have from time to time probs:
Number crunching : No progress change during calculation?

Where is the prob?

This are manufacturer OCed GTX260-216 55nm graphic cards. 4x EVGA SSC.
To now the other one, a GIGABYTE SOC (also GTX260-216 55nm) had no probs. But maybe because of too less WUs crunched.

I read now here, the old 65nm card have probs? Which kind of?

You can choose any driver before cuda3. Look at the suggested driver here.
http://developer.nvidia.com/object/cuda_archive.html

190 would be fine.

gdf

lkiller123, if you have a GTX260-192 you could test it with the CUDA 3.1 application on XP using the 258.96 driver, still available from NVidia (achieved drivers).

To run the CUDA 2.2 app on Windows you need a driver between 185.85 and 197.57 (inclusive).

To run the CUDA 3.1 app drivers from 257.21 through to 258.96 are required.

To run the CUDA 3.2 app (not presently in use, but soon to be tested) only the 260.89 or 260.99 drivers can be used.

http://www.nvidia.com/Download/Find.aspx

Use the Win7 64bit driver version. There are some differences between W7 and XP drivers but the app would really be what you are testing rather than the driver.

I'm still testing 260.99 but it's a messy picture, at least for the GT240's.

It seems to work on the Win7 x64 system and on one of two Vista systems, however on the Vista x64 system I have with 4 GPU's three of the cards are running at the correct speed, but one card is much slower, 405MHz rather than 579MHz and worse, the shaders on that card seem to be synced to the core at 810Mhz. The strange thing is that the card that dropped in frequency is identical to two other cards that did not drop in frequency.
I might try to install EVGA precision again but after I try a restart; I was able to change power management to maximum power, but it made no immediate impact.

It also looks like the 3.1 performance may not be as good as with 2.2 app for the 200series, but it will take a few more tasks to come in before I can be sure of that, and only for GT240’s.

For now I will leave the 260.99 on but if it's not as fast I may have go back to 197.57 - cant remember if I need dummy plugs for that one or not, but I have a spare omnicube.

In my view, after seen as Nvidia works, once a new architecture is out they don't care anymore about the older one. This happened for the g80 when the g200 came out, and to the g200 now that the fermi are out. The best driver is the one just before the change in GPU core: For g200 is the cuda2.2/2.3, and for fermi usually the latest.

That's why we give two applications out. We would like to support at least two generations, the G200 and Fermi now.

gdf

This is why I don’t think that the latest app and driver will fix the bug for the 60nm GTX260-192; they don't make a GTX260-192 any more.

After changing the power management to maximum power, I restarted the quad GT240 system and the card is now at the desired frequency (579MHz core and 1600MHz shaders). It has stayed at that for 2h 25min, so it looks good.

This GTX295 on XP x86 seems to be doing well with the 260.99 drivers; task completion times are the same for 3.1 as for 2.2, so the frequency can't be dropping off.

Betting Slip, you have only ran the slower IBUCH tasks since moving to CUDA 3.1. On my Fermi’s these are basically half speed tasks, using between 45% and 55% of the GPU. Wait until you have run a few faster WU's and you might want to test the new app when it comes out, just in case it turns out to be faster for 200series cards.

Sutaru Tsureku, you seem to have found a rare and endangered species - a working GTX260-192 (65nm)
one of its tasks

lkiller123, your clock speeds for both of those GTX260's is high, up from 1.24 to 1.51 and 1.49, and they are in the same rig. What are the GPU temperatures? All 3 failures came from device 1, device 0 has however not returned any completed tasks. Device 1 is at 1.49 GHz; is this the 65nm version? I would be happier to see those cards running in different systems. Putting a second card into a system usually increases the temperature of both cards, especially if you are crunching on your X4 940 as well. You might want to increase the fan speeds of both (I use EVGA Precision to do that).

All 4 of my GT240's (in the one rig) are still at the correct clocks, but it may be tomorrow before they return tasks that were running on a stable rig. So I will not be able to confirm performance differences of that Vista x64 system before then.
On my GTX470's the change in driver to 260.99 made no difference in terms of task performance; completion times were identical for same task types.

Anyone with Vista or Win7 and experiencing speed (frequency) problems should right click on the desktop, click NVidia Control Panel, on the left pane click Manage 3D Settings, on the right pane click the Program Settings Tab, make sure the Power Management Mode is set to Prefer Maximum Performance. If you had to change this you will need to restart the system to make it take effect.

On Win XP, there is no option to change from Adaptive Power Management to Prefer Maximum Performance. If anyone knows of an alternative method to force this please post it up. While this does not effect Fermi's (GTX470's at least) it does mean that some 200series cards go into power saving mode. Firmware updates might solve this, but just try getting the hold of one. I suspect the absence of a Maximum Performance option is a deliberate ploy by NVidia to undermine older generations of cards (they have done this in the past). It's like going into a garage to have your car checked over, and finding they topped up the windscreen washer reservoir but let two tyres down.

On my four GT240 card system I have 3 cards at 1599MHz and one at stock, 1340MHz. So I have to be careful to make sure I am comparing the same card.
The problem I have on comparing tasks on that system is that one of the cards dropped back to 405MHz, and I can't be sure the others did not at some stage. I also got 2 of the slower IBUCH tasks. I was not getting any of these slow IBUCH tasks with 6.05. Until now only Fermi's ran the 3.1 app, so the longer tasks are harder to finish in time for the full bonus. I expect you are right about the GT240's being slower, certainly unable to finish for the full bonus, but I just don’t have enough to go on to be sure about the performace; I have only completed 4 tasks. In a day or two I will know, and I want to wait and see how the 3.2 app works out for GT240's.

It looks like the cuda bug still effects your 65nm card, so you should stop crunching with it on GPUGrid. If the 3.2app fairs no better I think it is safe to say your 65nm card will never work here.

Sometimes a system restart can recover from that Windows error. Just keep the error message open and power down or restart the system.

Beyond, I would say I'm seeing a 10% drop in GPU utilization (Vista x64), but I would still like these tasks to finish before I compare them to my previous 6.05 tasks.

Not that we know. Apparently there's been sopme subtle hardware changes along with the die shrink.

MrS
____________
Scanning for our furry friends since Jan 2002

ACEMD beta 6.37 cuda2.2 for Windows and Linux is now out. Of course, these should be for g200 cards only and should not work for fermi.

Please accept beta work.

gdf

For GT240's the picture is simple; the CUDA3.1 app is slower than the CUDA2.2 app. and the 260.99 drivers are crap for XP. So people with GT240s should stay with the CUDA2.2 app, and not update to the latest drivers.

That ref. card was left at Adaptive power performance (now changed to Max). The reason for leaving it was to see if the frequency would drop or stay high; I wanted to get an idea of which cards dropped the frequency under which Operating System. On my four GT240 system one cards frequency dropped, but not the others. It also occasionally went back up to full performance only to drop again (a messy picture), and sometimes on a restart stayed high for a while (same as you observed). I wanted to work out if this was also happening on single card Vista systems and W7, but this is now irrelevant as it is clear that the CUDA3.1 app is simply too slow for GT240s.

Not sure why there are beta WUs out for CUDA2.2, thought tasks for the CUDA3.2 app were going to Beta instead, hence I kept the 260.99 drivers on? It will be tomorrow before I can go back to old drivers and get CUDA2.2 tasks again.

There was a problem with the server that it was distributing beta cuda2.2 to every cards as there was no cuda3 app. Now it's fixed.

gdf

Linux and Windows app seem to work well for cuda2.2.
So we can think of updating the official ones.

gdf

It was not a bug, it was fix to a bug as some linux users would hang the workunits.

gdf

Application 6.12 cuda2.2 for linux and windows are now out.

gdf

The are no unfriendly tasks. Beta work is only on for a few days before any release anyway.

gdf

How to disable recieving standard tasks that aren§t friendly to GTX260 65nm rev.A2 and get beta tasks only?
A few month ago it could be set, now I do not how to do it.

it's fine what you have.it should be the same speed.
gdf

sorry for off-topic, but:
right now i've got GTX275 running on 195.30 driver cudatoolkit 2.3 (linux). should I downgrade to 190.38 driver cudatoolkit 2.2? which one is faster - 2.3 or 2.2?

Due to the fact that cuda3.2 is still immature, we are coming out with the new application for cuda3.1 instead.
Tomorrow beta units.

gdf

ACEMDbeta 6.37 cuda3.1 for windows and linux are now available.
As usual you need to accept beta units.

gdf

How is it going with the cuda3.1 beta versions?

gdf
PS: This thread is for the new applications.

Re cuda3.1 beta versions,

i7-920, XP x86, Boinc 6.10.58, 2 x GTX470, 260.99:

04/11/2010 18:58:18 GPUGRID Scheduler request completed: got 0 new tasks
04/11/2010 18:58:18 GPUGRID Message from server: No work sent
04/11/2010 18:58:18 GPUGRID Message from server: No work is available for ACEMD beta version
04/11/2010 18:58:18 GPUGRID Message from server: Fermi-class GPU not supported by cuda2.2

- Just picking up normal (non-beta) work now.

@Sänger: well.. let's just say it's tradition. for most problems hard- and software go hand in hand, so in the end most stuff ended up being posted here. I guess that's why something "important for everyone" gets posted here rather than over there. Not exactly correct, though.

MrS
____________
Scanning for our furry friends since Jan 2002

If your p11-IBUCH_1_opt01_pYEEI_101027-3-20-RND2304 task is not running just abort it, but let it try to run for a few minutes and please post some details.

Thanks,

Looks like the techs moved the Betas to Live too soon, unless opt represents some sort of test WU.

Perhaps its more of an issue with IBUCH tasks than others, but I would suggest you abort it.

- I see you did. I'm worried that the next one will not be any better.

I am going to check it locally as well.

gdf

Sänger wrote:

Is the right one the acemd2_6.04_x86_64-pc-linux-gnu__cuda available at http://www.gpugrid.net/download/?

I don't know. The version I mentioned was for Windows.

Yet another example that the latest version is slower:

p16-IBUCH_11_PQpYEEIPI_101019-9-40-RND7287_0 2044054 4 Nov 2010 14:27:17 UTC 5 Nov 2010 08:56:36 UTC Completed and validated 66,265.58 21,205.96 7,645.29 11,467.93 ACEMD2: GPU molecular dynamics v6.12 (cuda)
p47-IBUCH_30_PQpYEEIPI_101019-9-40-RND4304_1 2033762 1 Nov 2010 22:56:28 UTC 2 Nov 2010 15:52:37 UTC Completed and validated 59,713.49 17,081.64 7,645.29 11,467.93 Not in DB

What about the cuda3.1 app in beta now?
gdf

So can I post BOINC tips here, or should I start a new thread? (Answer tomorrow, approaching bedtime here)

What about the cuda3.1 app in beta now?
gdf

It looks like I did not manage to pick up any of the 36 Betas.

We use to be able to select which ACEMD apps to crunch, and if I remember correctly, be able to crunch Betas only by unchecking both ACEMD apps.

Good luck,