GPU apps available for Einstein@Home's Gamma Ray Pulsar & Gravitational Wave searches

[Sunny129]

i apologize in advance for not having published these releases on our DC forum in a timely fashion...but the Einstein@Home project now has two new GPU applications for AMD, nVidia, and Intel GPUs - one for the Fermi LAT Gamma Ray Pulsar Search (FGRP3) and one for the Gravitational Wave Search (S6CasA). it appears the former was released back on 1/14/14 and the latter on 4/11/14. there are no specific requirements listed at the top of either of their respective threads over on the E@H message boards, and i could only find a handful of requirements specific to each app after having scoured both threads...these requirements are as follows:

• Fermi LAT Gamma Ray Pulsar Search (FGRP3) requires at least 1GB of VRAM (video memory)
• Gravitational Wave Search (S6CasA) appears to require 1GB of VRAM (video memory)...though i didn't see this stated definitively on the E@H message boards
• it appears only FP32 (single-precision) is required for the Fermi LAT Gamma Ray Pulsar Search (FGRP3)...but again, i didn't see this stated definitively on the E@H message boards
• your GPU must be FP64 (double-precision) capable for the Gravitational Wave Search (S6CasA) app
• for Gravitational Wave Search (S6CasA), you must set "Run beta/test application versions?" to "yes" in your Einstein@Home web preferences*

* it isn't entirely clear to me whether or not "Run only the selected applications - Gravitational Wave S6 Directed Search (CasA)" must be set to "yes" as well (in your Einstein@Home web preferences) in order to run these beta GPU tasks. intuition would tell me that yes it does, but i don't know for sure b/c i haven't tried to run the beta GPU apps with "Run only the selected applications - Gravitational Wave S6 Directed Search (CasA)" set to "no." for me, the only Einstein@Home tasks i'd like to run on this particular host are the Gravitational Wave beta GPU tasks...so i just set "Run beta/test application versions?" and "Run only the selected applications - Gravitational Wave S6 Directed Search (CasA)" to "yes," and set "use CPU" to "NO" to ensure i don't get any GW CPU tasks.

both of these apps are OpenCL coded for all GPUs, unlike Einstein@Home's two Binary Radio Pulsar Searches (BRP4G and BRP5), both of which have both OpenCL and Cuda coded apps (for AMD and nVidia GPUs respectively). happy crunching...

 

[Ken g6]

Thanks! Looks interesting.

Note that I had to turn on allowing beta test WUs to get these.

 

[VirtualLarry]

Interesting. I am currently running Einstein@Home on my Kabini dual-core AIO PC. It's an E1-2500. Does a Binary Pulsar whatever WU in around 24hrs.

I wonder how much VRAM is dedicated to the IGP, and if there's any way to increase that in order to process these new GPU WU types.

As an aside, I have no idea if the Kabini IGP (being GCN, I think) handles double-precision or not. Do the Binary Pulsar OpenCL WUs require DP?

Edit: Hmm, just checked my prefs, already have beta WUs enabled, still haven't seen any of these new ones. Maybe it's my (virtual) VRAM limitation.

 

[Rudy Toody]

Gamma Ray Pulsar Search took over 2 hours and pending is 74 credits.
Gravitational Wave Search errored out at approx 15 seconds.

I'm back on Milkyway for the 200K+ credits per day!

 

[networkman]

You may recall that I was a strong supporter of the Einstein@Home project crunching under the "Grand Rapids Radio Observatory" banner for quite some time, but after they ceased support for AMD/ATI video cards older than about the Radeon 4700 line, I quit the project and switched to others.

And like Rudy says, credit vs time is a serious factor to consider, especially when paying for your own electricity. Nevertheless, thanks for sharing Sunny129.

 

[Sunny129]

Thanks! Looks interesting.

Note that I had to turn on allowing beta test WUs to get these.

yes, i forgot to mention that part...usually all E@H beta tasks are run on the sister project Albert@Home as normal project tasks, and those apps don't get migrated to Einstein@Home until they're stable enough to be called release candidates...so this one is a little different. i'll append my OP to reflect those details.

As an aside, I have no idea if the Kabini IGP (being GCN, I think) handles double-precision or not. Do the Binary Pulsar OpenCL WUs require DP?

to my knowledge, neither the Binary Radio Pulsar OpenCL tasks nor the Fermi Gamma Ray Pulsar OpenCL tasks require double-precision - only the Gravitational Wave beta OpenCL tasks require double-precision.

I wonder how much VRAM is dedicated to the IGP, and if there's any way to increase that in order to process these new GPU WU types...

...Edit: Hmm, just checked my prefs, already have beta WUs enabled, still haven't seen any of these new ones. Maybe it's my (virtual) VRAM limitation.

i would imagine you might be able to manipulate the amount of dedicated VRAM in your mobo's BIOS. also, make sure that the "Run only the selected applications - Gravitational Wave S6 Directed Search (CasA)" is set to "yes."

 

[Assimilator1]

I'd imagine then that the strong DP ATI cards would do well on GW WUs .

This is another fascinating project that I've been wanting to do since before it was released! :$

 

[Sunny129]

to give you an idea Mark, E@H GW tasks take approx. 10 hours to complete on my AMD Phenom II x6 1055T CPU @ 3.3GHz (running 6 tasks simultaneously), whereas they only take approx. 20 minutes to complete on my HD 7970 (running 3 tasks simultaneously). if you do the math, that's a 15-fold increase in performance over the CPU version of the GW app. i'm sure my HD 7970 is capable of cutting those run times down substantially if i weren't loading up the remaining CPU cores with CPU tasks from other projects...

 

[Assimilator1]

Nice! , my 5850 should do reasonably well then, thinking about it it did a single task yesterday when I set it up........

Yes, but a Gamma-ray pulsar search task, 6341s run time & 3681s CPU time, why the high CPU time? I did snooze the GPU for an hr, but surely that won't affect the CPU time?

 

[Sunny129]

Yes, but a Gamma-ray pulsar search task, 6341s run time & 3681s CPU time, why the high CPU time? I did snooze the GPU for an hr, but surely that won't affect the CPU time?

while the GPU version of the Fermi Gamma Ray Pulsar Search app is no longer considered beta, it is still in its infancy. per project administrator and developer Bernd Machenschalk:

- OpenCL versions for AMD, NVidia and Intel GPUs. As we did with the first BRP GPU apps, so far only the FFT will run on the GPU, the rest of the computation (which is still a lot) is done on the CPU. As the application develops, more and more computation will be shifted to the GPU. Note that in the project preferences setting there is (now) a "GPU utilization factor for FGRP" separate from the BRP one.

so the high CPU times exist simply b/c a substantial chunk of each FGRP3 GPU task is still crunched on the CPU. as the application matures, more and more of it will run on the GPU, gradually lowering the fraction of CPU time required for each GPU task.

 

[Assimilator1]

Ah I see, makes sense now, thx .