>if you dont like my 9500GT and 8800GTS for Cuda
And I'd really like to use my Z80 for BOINC too, but it won't let me. If you don't have the correct equipment... :)

---

Cheers,

PeterV

.

When my CPU (i7) is working full with other BOINC projects, calculation time for Milkyway/CUDA is 9.5 minutes. When my CPU is idle, calculation time is 5.5 minutes. Close to two times difference !

Edit: confirmed on my second computer.

The GTX 285 that I have takes about 5:24 per WU with no load on the CPU and leaving the computer idle, this is with the WU that begin with gs_s222_3s.

@Anthony:

I sent you a solution to this problem by PM!

I just think at this point we should just be happy to have the app running. CP has been of great help and everyone involved in this project. Granted, the timeline somewhat took a long time but who am I. I am not a scientist or a coder. I am just happy to crunch MW on my GTX285. It will improve I am sure.

Thanks to all of you who made this possible. It is greatly appreciated even though I never voiced it but I just did.

Martin Chartrand

When my CPU (i7) is working full with other BOINC projects, calculation time for Milkyway/CUDA is 9.5 minutes. When my CPU is idle, calculation time is 5.5 minutes. Close to two times difference !

Edit: confirmed on my second computer.

The GTX 285 that I have takes about 5:24 per WU with no load on the CPU and leaving the computer idle, this is with the WU that begin with gs_s222_3s.

@Anthony:

I sent you a solution to this problem by PM!

Thanks a lot, I'm implementing it now so the new version should be up within the hour.

Edit: The newest version is now on BOINC, the completion times should be more stable around 5-6 minutes, thanks to Cluster Physik.

Ok, after few restarts it works!!!!

But I install driver 190.62, BOINC 6.10.1 and detached + atached to projest and set only GPU...

It works witch AQUA@HOME CPU.

Win7 64 bit.

On GTX260 (625Mhz GPU 1348 Shader) = 11 m. 25 s. = 685s. 74.24 credits so 1 credit in 9.23s.

GPUGRID pays better..... (1 credit at 4-7sec.)

But MW app is "young" and I hope soon it will be improved to have better performance :)

GPUGRID uses single precision math which is 8 times faster than the double precision math that Milkyway@Home requires. For reference, the single precision version of the Milkyway@Home CUDA app can do a WU in 15 seconds, while the double precision version takes upwards around 85 seconds. Granted the WU were smaller than the ones on BOINC, but it clearly shows the performance between single and double precision.

the new coming nVIDIA G300 Cards will calculate MW faster (faster double prec. support)

the Q about single prec in MW: ATI users would like it too with 4s / WU ?

the current nVIDIA cards are to slow in double prec. - and the credits are high enough if you got the "best" MW card - an ATI 4870 will get over 75credits per minute (1 WU) or over 108000 Credits / day - this is enough

the new coming nVIDIA G300 Cards will calculate MW faster (faster double prec. support)

the Q about single prec in MW: ATI users would like it too with 4s / WU ?

the current nVIDIA cards are to slow in double prec. - and the credits are high enough if you got the "best" MW card - an ATI 4870 will get over 75credits per minute (1 WU) or over 108000 Credits / day - this is enough

Due to the nature of our work here and lots of discussions with our astronomers, we can't support single precision GPU work because it's not precise enough.

We're doing calculations over very large (I dare say astronomical) distances, and single precision work simply doesn't have the accuracy to correctly model our data.

---

> this is enough
ENOUGH?!
Its never enough. :)

---

Cheers,

PeterV

.

GPUGRID uses single precision math which is 8 times faster than the double precision math that Milkyway@Home requires. For reference, the single precision version of the Milkyway@Home CUDA app can do a WU in 15 seconds, while the double precision version takes upwards around 85 seconds. Granted the WU were smaller than the ones on BOINC, but it clearly shows the performance between single and double precision.

Thanks a lot, I'm implementing it now so the new version should be up within the hour.

Edit: The newest version is now on BOINC, the completion times should be more stable around 5-6 minutes, thanks to Cluster Physik.

Great Job!

Now MW WU (with AQUA@HOME CPU on 4 cores) takes 5m 57s - great improvement :)

So: 1 credit takes 4,85 sec. and now MW is as good as GPUGRID :)

If with next version of program you will go down below four minutes you will be a CodeMaster :)

---

A proud member of the Polish National Team

COME VISIT US at Polish National Team FORUM

@Anthony:
I sent you a solution to this problem by PM!

Thanks a lot, I'm implementing it now so the new version should be up within the hour.

Edit: The newest version is now on BOINC, the completion times should be more stable around 5-6 minutes, thanks to Cluster Physik.

Great Job!

Now MW WU (with AQUA@HOME CPU on 4 cores) takes 5m 57s - great improvement :)

So: 1 credit takes 4,85 sec. and now MW is as good as GPUGRID :)

If with next version of program you will go down below four minutes you will be a CodeMaster :)

My current guesstimate is that a GTX280 could do a WU in about 3 minutes. With some more tweaks to the code maybe even slightly faster. When I find the time to test those tweaks (they are likely going into the ATI app too, about 10% improvement can be expected) I will send them Anthony. But don't expect 10% higher credits per second from those changes as the number of executed operations is going down by roughly the same amount. That means the credits per WU needs to be corrected for this.

OK -- so if I understand this correctly, for this project, GPU support will be limited to what are essentially high end cards (due to the double precision requirement).

That list for ATI GPU's is -- 38xx, 47xx??, 48xx. The list for nVidia is - GTS 260 and up (excluding any of the 98xx series including GTX, as well as the GTS 250 (largely a reconfigured 9800GTX with somewhat lower power requirements).

My understanding in terms of what that means cost wise is that aside from some of the more cost efficient ATI cards (running just under $100), the requirement is in the >$150 range, and further that the requirement includes solid power supplies -- perhaps 550W (REAL rather than cheap), as well as really solid cooling.

Fair enough, I was hoping a GTS 250 I picked up would work, but I guess it will live its life in GPUGrid land instead.

---

Sorry to cross post this from another thread but maybe it can be answered here, as nothing heard from the other thread.

http://milkyway.cs.rpi.edu/milkyway/forum_thread.php?id=631&nowrap=true#29971

---

Proud Founder and member of



Have a look at my WebCam

I noticed that no progress bar is showing with the new application version from Aug 30. Everythings else - credit, runtime etc - is the same as with the old version.

Anyone else noticing the same behaviour?

I noticed that no progress bar is showing with the new application version from Aug 30. Everythings else - credit, runtime etc - is the same as with the old version.

Anyone else noticing the same behaviour?

Yes, I've uploaded a new version about an hour ago to fix the issue, thanks for reporting this.

GPUGRID uses single precision math which is 8 times faster than the double precision math that Milkyway@Home requires.

With the difference being that big even in hardware, I wonder if it could be worth emulating double precision in software.. I know the maths to do it exists, although I've forgotten how it works.

WOO HOO! Hurray the great day has come! Congrats on getting the cuda app going. As of this morning running MW on 1nvidia+1ati..lol :)
Got pleanty road ahead but good job to everyone (on all fronts) getting this multiproject multi gpu thing boincing ;)

Attached the c1060 (having set web prefs to gpu) it downloaded the app and away it went..
92312897
92312896

..I can confirm others reports. Current draw shows GPU running ~45-55% /w MW cuda. Also, it would be nice to see elapsed time in stderr_txt?

Keep crunching & keep up the good work MW!

---

(i7 with 4x GTX295, 8 GPUs) Thank you for developing this CUDA version !

I notice that the CUDA app almost completely freezes the system though. It would really help if there was more screen updating inbetween, at least for the main video adapter. In the current setting it took me an hour to regain control over the machine, after yanking the network cable to make it run out of workunits.

I also notice that Boinc manager has a hard time communicating with the apps, it almost constantly shows a popup "Communicating with BOINC client, please wait". this prohibits control over Boinc. in case of a problem one cannot stop the application gracefully.

I also notice that Boinc manager has a hard time communicating with the apps, it almost constantly shows a popup "Communicating with BOINC client, please wait". this prohibits control over Boinc. in case of a problem one cannot stop the application gracefully.

I had that problem with version 6.6.38, regardless of the type of app.. made it completely unusable. Don't know if that helps at all, just throwing it out there :P

This is where one of the new features in 6.10.x comes in handy.

It has a button to only show active tasks only.

---

Well, the new button will not do any good, it doesn't even get to showing buttons most of the time.

And this version of boinc worked just fine with GPUGrid, so it's the app's way of minimizing communication. 'Maximizing runtime on the GPU' is the positive approach to describing the same problem.

FTP access works just fine, so it is the video updating that is the bottleneck.

Thanks for you input though.

---

Team Bletchley Park.