NVIDIA: "Deprecation plans for 32-bit Linux-x86 CUDA toolkit and CUDA driver"
13 Nov 2013 9:28:07 UTC
Topic 197270
(moderation:
FYI, NVIDIA recently made this announcement :
The essence is that for the moment, X86 32 bit development kits and drivers on Linux are still supported, but from now on NVIDIA might stop supporting this platform at any time.
We still get a reasonably large share of results from 32 bit CUDA Linux installations, so there are no immediate plans to drop 32 bit CUDA Linux support on E@H. But we will periodically evaluate the situation (also for other OSes, e.g. Windows 32 bit is beginning to look like an endangered species...).
Cheers
HB
NVIDIA: "Deprecation plans for 32-bit Linux-x86 CUDA toolkit and
)
Thanks for the heads-up; hope it won't be too soon as everything we run is 32-bit linux (and will be for some time to come).
I'd guess there is quite a
)
I'd guess there is quite a bit of 32-bit Windows XP running in corporate shops which skipped the Vista step, and liked the result so well that they kept on skipping. Most of those may be scared into upgrade by the announced end of XP support (April 8, 2014), but I'd bet there will be still a long, slow tail of personal usage. If Win8 had been better received in the corporate world this would go easier.
But you have access to the numbers on whether their contribution Einstein is worthwhile.
I, personally, run two 32-bit XP hosts on Einstein. One is slated for retirement just as soon as I finally finish fiddling with the replacement, which I built over a year ago but have not gotten around to final configuring for the desired use.
The second I vaguely intend to replace with a new build I vaguely intend to start early in 2014. So I am aiming to be off XP around the currently claimed time of support termination.
I'll place a small bet that the termination date will get pushed out a bit.
The announcement from NVidia
)
The announcement from NVidia applies to Linux only. I guess they will support 32Bit Windows a lot longer.
BM
BM
RE: The announcement from
)
My understanding is that is only for their CUDA support. The 32-bit GPU drivers for Linux will be around for some time yet.
And then again... If you are on Linux, why on earth are you still on 32-bits?!
(OK, so I'll admit to still running old 32-bit Intel Atoms and even an old Epia Via C3 32-bit board. Then also, they are not running any GPUs... :-P )
Keep searchin',
Martin
See new freedom: Mageia Linux
Take a look for yourself: Linux Format
The Future is what We all make IT (GPLv3)
A quick check on a new boinc
)
A quick check on a new boinc install on Linux shows in the boinc home directory:
libcudart.so -> /opt/cuda/lib/libcudart.so
which is a 32-bit library.
There is also /opt/cuda/lib64/libcudart.so, so why has the boinc installer kept to only 32-bits?
How do boinc tasks pick up what libcudart.so to use from where?
Happy crunchin',
Martin
See new freedom: Mageia Linux
Take a look for yourself: Linux Format
The Future is what We all make IT (GPLv3)
What you wrote only affects
)
What you wrote only affects the BOINC client, not necessarily project apps ("tasks"). Einstein@Home CUDA apps copy their own CUDA runtime library (libcudart) into the slots directory where the task is executed, I would every other project expect to do the same.
BM
BM
RE: And then again... If
)
No point in suffering the (admittedly usually small) 64-bit overhead unless running apps that need more than 4Gb *per process*, or have other special needs where 64-bit is useful. And, of course, einstein@home is 32-bit - which is the most computationally intense thing my computers do.
Find the binary for the task, and do 'ldd $NAME'. I'm not familiar with einstein binaries per se, but generally the library path can be built into the binary, appear somewhere in ld.so.conf, or overridden through environment variables.
RE: No point in suffering
)
In x64 double the amount of registers is available - which leads to an admittedly usually small performance improvement.
MrS
Scanning for our furry friends since Jan 2002
Indeed, whether the extra
)
Indeed, whether the extra registers would help e@h more than the 64-bit overhead would cost is an open question.
Still, with linux at least you get the choice; 64-bit kernels work perfectly with 32-bit userspace, while still running 64-bit applications natively.
Back to the original note, it does seem strange that nvidia would drop CUDA support for 32-bit linux and retain it for Windows; possibly it's a marketing move, or preparing the ground for a similar withdrawl on windows.
RE: I'd guess there is
)