BRP CUDA requirements ( so you can view in the one place )
30 Dec 2010 5:18:38 UTC
Topic 195545
(moderation:
The new BRP CUDA tasks require the following:
- Windows
- BOINC 6.10.x
- CUDA driver >= 3.2 (>= 260.00)
- 100% GPU (it just uses up to 75%)
- 20% CPU
- 250 MB RAM required (~210 used)
- Speed up compared to CPU: up to 20x (240 cores)
courtesy of Bernd and Oliver, via Ageless. :-)
Cheers, Mike.
I have made this letter longer than usual because I lack the time to make it shorter ...
... and my other CPU is a Ryzen 5950X :-) Blaise Pascal
BRP CUDA requirements ( so you can view in the one place )
)
Daer Mike,
thanks for the complete list of BRP requirements.
Just one small addition: From the ration CPU time vs.
wall time it seems to take 50% of one of my Intel
Q6600 cores to feed the 336 cores of my slightly
overclocked Gigabyte GTX460.
Happy 2011 and may the gravitational waves be with you ;-)
Armin
RE: Daer Mike, thanks for
)
Noted. :-)
I suppose if anyone else discovers other significant variations from our quoted spec then please simply post here your observations ( but kindly seek elsewhere for troubleshooting - even start your own thread if need be ). The dev's have their test rigs, but of course us punters dwarf them for permutations of hardware.
And you! Be safe. :-)
Cheers, Mike.
I have made this letter longer than usual because I lack the time to make it shorter ...
... and my other CPU is a Ryzen 5950X :-) Blaise Pascal
Changes. As per
)
Changes. As per here:
The inevitable question: when
)
The inevitable question: when will a linux app come ? And when we are at it: how big is the part of linux in the total crunching ?
We get about 60% Windows, 30%
)
We get about 60% Windows, 30% Linux and 10% Mac OS X results for S5GC1HF.
The numbers are not very stable for BRP3, since we still keep changing things there. But I don't think that a comparably large fraction of Linux hosts attached to Einstein@home is capable of runnning our CUDA App as there is for Windows. Consequently the Linux CUDA App doesn't have a priority as high as the Windows one.
BM
BM
RE: We get about 60%
)
You've got at least one GTX-400-series Linux cruncher here available for e@h...
I would expect the Linux Boinc enthusiasts to have a higher proportion of cutting edge hardware than for Windows... (A summary comparison for e@h would be very interesting. ;-) )
Note that 'numbers of users' is of little consequence. It's more the credits counts that indicate the relative returns for the project. A very few users may well be the one's disproportionately contributing the most credits...
Happy fast crunchin',
Martin
See new freedom: Mageia Linux
Take a look for yourself: Linux Format
The Future is what We all make IT (GPLv3)
ML1: "Cutting edge hardware"
)
ML1:
"Cutting edge hardware" is often not the right choice for Linux, because the driver availability is not as good as for Windows yet. Therefore it is a good choice not to use the most modern hardware for a Linux desktop.
Nevertheless it`s more comfortable to run Cuda on Linux (for my own: Ubuntu 10.10) than on Windows 64-bit. On Windows you first have to login as system administrator so that BOINC starts to work. Hereafter you have to change to a simple user account without leaving the admin-account. The security system of Windows 64-bit doesn`t allow a simple user to run BOINC-cuda.
By the way:
What is the best NVIDIA-GPU for CUDA running einstein@home? Does the application only use up to 240 GPU-cores, so that more are useless? Which graphic cards were used as reference cards while developing the new CUDA-application?
Intel Q9300 Quadcore, 2500 Mhz, 4096 MB RAM, GeForce 9800 GT, Vista Ultimate 64-bit, Ubuntu 10.10 64 bit
RE: Does the application
)
I don't think so. At least the cuFFT should use whatever is there.
For E@H we're developing on Tesla C1060, our Windows test machines usually have GeForce GTX 285.
BM
BM
Stderr output A lot of
)
Stderr output
A lot of workunits have finished with the following error message on Ubuntu-Linux 10.10:
6.10.56
Eine symbolische Verknuepfung kann nicht in einem Registrierungsschluessel erstellt werden, der bereits Unterschluessel oder Werte enthaelt. (0x3fc) - exit code 1020 (0x3fc)
Activated exception handling...
[05:15:55][5876][INFO ] Starting data processing...
[05:15:55][5876][ERROR] Couldn't initialize CUDA driver API (error: 100)!
[05:15:55][5876][ERROR] Demodulation failed (error: 1020)!
05:15:55 (5876): called boinc_finish
]]>
What is the reason for that?
Intel Q9300 Quadcore, 2500 Mhz, 4096 MB RAM, GeForce 9800 GT, Vista Ultimate 64-bit, Ubuntu 10.10 64 bit
RE: ML1: "Cutting edge
)
You'd be surprised... I certainly am at how now most recent hardware is automatically supported. Both Intel and AMD are serious about 'Open Source' support... And others. It's always worth a check!
If you're cautious, then hold off for about 6 months to allow for any updates to have rippled through. Then again, for such as nVidia graphics cards, the drivers can be updated for Linux before whatever cards even become available!
I started out on 8-bit micros, moved up to unix, then moved over to 16/32-bit Windows, then various shortcomings in WinXP pushed me over to Linux (on 32-bit hardware). Once I'd caught up with the Linux development, I was left wondering why I'd not moved over sooner... Meanwhile, even the latest Windows still appears to be a single user system with multi-user profiles bolted on top even though the Windows kernel could be used in a true multi-user way...
Which comes to... Why are you running a 32-bit version of Linux on 64-bit hardware? You may as well take advantage of the 64-bits.
Happy fast crunchin',
Martin
(On Mandriva 2010.2 64bit, but not recommended for first time users! Take a look at Kubuntu or Ubuntu for a first try.)
See new freedom: Mageia Linux
Take a look for yourself: Linux Format
The Future is what We all make IT (GPLv3)