I'd like to refresh the subject of CUDA 5.5 apps. CUDA 5.5 platform has been issued by NVIDIA in August, and CUDA 6.0 is on the horizon.
You said a while ago that CUDA 5.5 app would see the light of day only if it is fast enough. Can you shed some light on this subject? :)
I'd like to refresh the subject of CUDA 5.5 apps. CUDA 5.5 platform has been issued by NVIDIA in August, and CUDA 6.0 is on the horizon.
You said a while ago that CUDA 5.5 app would see the light of day only if it is fast enough. Can you shed some light on this subject? :)
Internal testing of compatibility and performance has been completed. The gain of CUDA 5.5 apps compared to our CUDA 3.2 apps is rather mediocre (about 15%). Still we intend to publish CUDA 5.5 apps in the foreseeable future. Setting up an automatic build environment for such apps is next in the pipeline and being done (well, not during holidays of course), however not with highest priority.
Fifteen percent gain in speed for CUDA5.5 comp. to 3.2 sounds good to me (I guess it'll vary much between different generations of NVIDIA GPUs).
Can you give us estimated time of arrival for new apps?
Fifteen percent gain in speed for CUDA5.5 comp. to 3.2 sounds good to me (I guess it'll vary much between different generations of NVIDIA GPUs).
Can you give us estimated time of arrival for new apps?
Luke
Work at the institute will resume on the 6th of Jan for the E@H team. A CUDA 5.5 app will first be tested at Albert@Home for some time, so if all goes well it could be in production in late Jan or early Feb 2014 I guess, as a conservative estimate. At the moment our priority is a brand new FGRP (Fermi gamma ray pulsar) search (better data, much improved search algorithm, GPGPU support ....) which is now on Albert.
For BRP, there's some potential for improvement. I'd like to cut down the amount of data that has to be sent over the PCIe bus, and we need to check out some alternative for the FFT part in OpenCL.
So perhaps for CUDA we can get a 20% improvement on Kepler cards, if all oes well.
We did 57 and 58 some time ago (also 59, and 69-72).
I only know this from processing the joblog file, and wondered what was going to happen when the PA series got up to the range we've seen before. The answer is, clearly it's all under control!
Server status says now less than a month work left. Is there more work to be expected on BRP5 after that? I'm happy with the available other GPU work as well. Just curious.
Server status says now less than a month work left. Is there more work to be expected on BRP5 after that?
Yes. Essentially we will extend the parameter space, i.e. search the same data again with a different template bank. This second half of BRP5 should last about as long as BRP5 lasted so far.
Fifteen percent gain in speed for CUDA5.5 comp. to 3.2 sounds good to me (I guess it'll vary much between different generations of NVIDIA GPUs).
Can you give us estimated time of arrival for new apps?
Luke
Work at the institute will resume on the 6th of Jan for the E@H team. A CUDA 5.5 app will first be tested at Albert@Home for some time, so if all goes well it could be in production in late Jan or early Feb 2014 I guess, as a conservative estimate. At the moment our priority is a brand new FGRP (Fermi gamma ray pulsar) search (better data, much improved search algorithm, GPGPU support ....) which is now on Albert.
For BRP, there's some potential for improvement. I'd like to cut down the amount of data that has to be sent over the PCIe bus, and we need to check out some alternative for the FFT part in OpenCL.
So perhaps for CUDA we can get a 20% improvement on Kepler cards, if all oes well.
Hello Bernd! I'd like to
)
Hello Bernd!
I'd like to refresh the subject of CUDA 5.5 apps. CUDA 5.5 platform has been issued by NVIDIA in August, and CUDA 6.0 is on the horizon.
You said a while ago that CUDA 5.5 app would see the light of day only if it is fast enough. Can you shed some light on this subject? :)
Regards!
Luke
RE: Hello Bernd! I'd like
)
+1
-----
Internal testing of
)
Internal testing of compatibility and performance has been completed. The gain of CUDA 5.5 apps compared to our CUDA 3.2 apps is rather mediocre (about 15%). Still we intend to publish CUDA 5.5 apps in the foreseeable future. Setting up an automatic build environment for such apps is next in the pipeline and being done (well, not during holidays of course), however not with highest priority.
BM
BM
Thanks Bernd! Fifteen
)
Thanks Bernd!
Fifteen percent gain in speed for CUDA5.5 comp. to 3.2 sounds good to me (I guess it'll vary much between different generations of NVIDIA GPUs).
Can you give us estimated time of arrival for new apps?
Luke
RE: Thanks Bernd! Fifteen
)
Work at the institute will resume on the 6th of Jan for the E@H team. A CUDA 5.5 app will first be tested at Albert@Home for some time, so if all goes well it could be in production in late Jan or early Feb 2014 I guess, as a conservative estimate. At the moment our priority is a brand new FGRP (Fermi gamma ray pulsar) search (better data, much improved search algorithm, GPGPU support ....) which is now on Albert.
For BRP, there's some potential for improvement. I'd like to cut down the amount of data that has to be sent over the PCIe bus, and we need to check out some alternative for the FFT part in OpenCL.
So perhaps for CUDA we can get a 20% improvement on Kepler cards, if all oes well.
Cheers
HB
This morning I got PA59.
)
This morning I got PA59. Before that PA56. Are we skipping 57 and 58? Or due to higher demand because of the BRP4G which ran out?
We did 57 and 58 some time
)
We did 57 and 58 some time ago (also 59, and 69-72).
I only know this from processing the joblog file, and wondered what was going to happen when the PA series got up to the range we've seen before. The answer is, clearly it's all under control!
Server status says now less
)
Server status says now less than a month work left. Is there more work to be expected on BRP5 after that? I'm happy with the available other GPU work as well. Just curious.
RE: Server status says now
)
Yes. Essentially we will extend the parameter space, i.e. search the same data again with a different template bank. This second half of BRP5 should last about as long as BRP5 lasted so far.
BM
BM
RE: RE: Thanks
)
Hi,
Cuda 5.5 is still not there. Shure, I know, there were many many problems. But:
http://einstein.phys.uwm.edu/forum_reply.php?thread=10028&post=128444#input
So maybe it makes sense to skip cuda 5.5 and switch directly to cuda 6 ?