Projects that support Intel HD GPU crunching

[zombie67]

Ding!

I am now crunching AstroPulse tasks (SETI@home) on my win8 Intel HD 4000 (i7-3770S CPU) (http://setiathome.berkeley.edu/show_host_detail.php?hostid=6937166). They take 10,300 seconds to run. I am *slowly* downloading some tasks for my 7970, which will be run with the lunatics .40 app, so I can compare. Stay tuned. Also, I am not sure how that compares to other GPUs, or what my GPU load is. Does anyone know of an afterburner alternative that "sees" Intel HD GPUs?

See here (http://setiathome.berkeley.edu/forum_thread.php?id=70717) for info on how. Note my post where I have a modified app_info.xml, that tricks the server into sending tasks to my CPU, because the SETI server has not yet been upgraded to "know" about the Intel HD GPU. Yes, you read that right...S@H server code is not up to date. :mad: The good news is that the tasks are the same for CPU and GPU, so my GPU app crunches those "CPU" tasks just fine.

Things to consider before you try this:


I have not yet tried to do this with a machine that *also* has a discrete GPU. I think I have read something about problems where the opencl driver for intel conflicts with the opencl driver of amd/nvidia.
I have not tried this with win7. I have read conflicting reports on it working with win7. For sure nothing less than win7.
You have to run something later than the "recommended" version of BOINC. I am using .52 for this machine.
You won't be able to also run CPU versions of AP at the same time, if that matters to you.
AP tasks are BIG, so you really need to apply the TCP RFC1323 extensions, if you want them to download....well ever. See more here (http://setiathome.berkeley.edu/forum_thread.php?id=71002). Turns out that the Linux and OSX TCP implementation already include those features. It's just win that is lacking by default.



Next up: I hear Slicker is working on the app and server at collatz. Fingers crossed!


Edit #1: 7970, running a single task with v40 lunatics takes 2,300 seconds.

Edit #2: GPU-Z supports monitoring GPU load on Intel HD

[Fire$torm]

Thanks Z for making this effort! :-bd

Btw, I took the liberty of making this thread a sticky. Hope you don't mind.

F$

[zombie67]

I have PG PPS Sieve running on my Intel HD 4000, using the ATI openCL app. It requires faking out the Server with app_info.xml. The admin said they would update the server to work natively, if I could get it working. So I guess the ball is in their court now. But you can run it using app_info.xml until then.

Details here:

http://www.primegrid.com/forum_thread.php?id=5136

[Fire$torm]

Damn Z,

What the heck... You had a real fight on your hands in that thread... Jeeeeeez, Whats up with people these days...?

Anyhoot, GJ getting your HD 4000 to crunch on PG. :-bd

[Slicker]

Collatz now supports Intel GPU apps w/o an app_info.xml file for the solo_collatz app. Several volunteers are helping with getting the app tuned so it uses as little CPU w/o reducing GPU load too much.

[zombie67]

Also, the SETI Beta server has been upgraded to "know" intel GPUs, and they added the intel GPU to the preferences page. Why are those two different actions required by the admin? Anyway, if you want to run astropulse there, with your intel HD 4000 or better, go for it.

[Slicker]

Also, the SETI Beta server has been upgraded to "know" intel GPUs, and they added the intel GPU to the preferences page. Why are those two different actions required by the admin? Anyway, if you want to run astropulse there, with your intel HD 4000 or better, go for it.

Because there was no real design when coprocessors were implemented because it was assumed that NVidia would be the only one. Anything added later was really a hack. In fact, even the NVidia code is a hack if you ask me. The data, what little there is, is stored in the "serialnum" field of the database of the server. Half the data isn't stored at all but is sent over and over again with each work request which more than doubles the amount of data sent and received. Stupid. GPUs are synonomous with serial numbers, right? Not in my universe. Anyway, not only does the admin need to change the code in the sched_customize.cpp so it will send work, he also needs to change the project include files so the php code can set the GPU preferences. Why? Because the preferences aren't in their own table. They also are ALL stored in a single database variable as an XMLish string.

I don't know how a certain someone, who has a PhD in IT, can approve such a design unless he flunked his database design class as an undergrad. When I used to teach database design at Triton Community College, I can tell you that any design which stored multiple data points in a single database field got the student an F on the assignment since that is just being lazy.

[zombie67]

Albert how supports it:

https://albert.phys.uwm.edu/forum_thread.php?id=8954&nowrap=true#112627

[zombie67]

Einstein now supports intel GPU too. And without having to do any funky app_info.xml stuff. Just turn it on in your preferences.

[Slicker]

Einstein now supports intel GPU too. And without having to do any funky app_info.xml stuff. Just turn it on in your preferences.

From the BOINC Dev email list, it sounds like using an app_info with the Intel GPU app still isn't working but they are working on it. Hopefully now that the server side is working, projects won't need users to use an app_info other than for testing new apps.

[DrBackJack]

That is wicked running on HD CPU like that. On top of that you can run other GPU tasks on your video card at the same time ?

[zombie67]

That is wicked running on HD CPU like that. On top of that you can run other GPU tasks on your video card at the same time ?

Yep! All the normal task(s) on your GPU cards, PLUS tasks on your intel HD graphics.

[zombie67]

I now have POEM running on Intel HD 4000 with app_info.xml. Files here:

https://dl.dropboxusercontent.com/u/55884901/POEM_Intel.zip

[zombie67]

SETI & SETI Beta now support intel GPUs natively. No xml files required.

http://setiathome.berkeley.edu/forum_thread.php?id=73935

[Maxwell]

Sorry to revive a dead thread, but I just noticed that I have some Intel HD graphics in my chips. Specifically, HD 3000 in my i7-2600k (http://www.newegg.com/Product/Product.aspx?Item=N82E16819115070). After reading this thread and a few of the project threads discussing Intel HD crunching, I have a little hope for my chip. Am I correct in assuming that no project will run on my HD 3000?

[zombie67]

HD 4000 or greater is required. HD 3000 will not crunch.

[Maxwell]

HD 4000 or greater is required. HD 3000 will not crunch.
Thanks! That's what I needed to know.

[denim]

Out of nowhere, my office machine started crunching SETI v7 7.03 on the IntelHD graphics. Very odd as I have always had the IntelHD box checked on the profile settings and this machine had not crunched it in the past, same goes for Einstein. Either way, I will take it as I really need to help our team namesake a little better and get my team project rank up a bit.

[FourOh]

Out of nowhere, my office machine started crunching SETI v7 7.03 on the IntelHD graphics. Very odd as I have always had the IntelHD box checked on the profile settings and this machine had not crunched it in the past, same goes for Einstein. Either way, I will take it as I really need to help our team namesake a little better and get my team project rank up a bit.

I just checked out the SETI apps page, I guess I didn't realize Astropulse v7 was a new app as of a few months ago. At some point I suppose I'll run it for the hours... but if I recall Astropulse work is intermittent. I'm not ready for that headache just yet!

[Maxwell]

I just checked out the SETI apps page, I guess I didn't realize Astropulse v7 was a new app as of a few months ago. At some point I suppose I'll run it for the hours... but if I recall Astropulse work is intermittent. I'm not ready for that headache just yet!
Backup projects. On the GPU on which i'm running SETI, I have PG set up as a back project. That way on the frequent occurrences when I run out of Astropulse work, it's still working on something.

[zombie67]

Collatz also works a a backup for intel HD.

[Maxwell]

Collatz also works a a backup for intel HD.
Yes - I forgot what thread I was in. I use Collatz as a back for Intel HD crunching. Which worked flawlessly when I ran out of Einstein WUs to crunch on the Intel GPU over the weekend...

[FourOh]

Collatz also works a a backup for intel HD.

Lately I haven't been running any Intel HD apps because Einstein and Collatz both want at least half a CPU thread to go along with the GPU tasks, and I've been trying to meet some CPU credit goals. What kind of CPU commitment do the SETI Intel HD apps require? Anyone know off hand?

[denim]

Lately I haven't been running any Intel HD apps because Einstein and Collatz both want at least half a CPU thread to go along with the GPU tasks, and I've been trying to meet some CPU credit goals. What kind of CPU commitment do the SETI Intel HD apps require? Anyone know off hand?

Right now, mine is saying 0.00042% CPU + 1 Intel GPU.

[denim]

The Einstein WU's are averaging less than 0.20 (yes you read that right) claimed credit.

[denim]

Where as SETI is granting 42.0 credits per WU for almost the same crunching time.

[FourOh]

The Einstein WU's are averaging less than 0.20 (yes you read that right) claimed credit.


Where as SETI is granting 42.0 credits per WU for almost the same crunching time.

Einstein grants 62.5 for an Intel GPU BRPS work unit, regardless of what is claimed. Daily production on HD 4xxx is anywhere from 7,000 to 10,000 credits.

[Maxwell]

The Einstein WU's are averaging less than 0.20 (yes you read that right) claimed credit.
Claimed credit is based off of CPU time, which is pretty minimal with well-programmed GPU apps. The Einstein WUs are granting 62.5 credits each.

[denim]

Claimed credit is based off of CPU time, which is pretty minimal with well-programmed GPU apps. The Einstein WUs are granting 62.5 credits each.

I refreshed the page after and noticed the difference. So I was wrong and it appears that Einstein has the better return of the two. Either way, I am keeping the taps open for both to keep being crunched on this one machine.

[denim]

Einstein grants 62.5 for an Intel GPU BRPS work unit, regardless of what is claimed. Daily production on HD 4xxx is anywhere from 7,000 to 10,000 credits.

That is a crap load more than I will be able to produce. If I get 1000 from either project per day, I will be happy.

[Maxwell]

That is a crap load more than I will be able to produce. If I get 1000 from either project per day, I will be happy.
Mine gets ~5k/day on Einstein running two WUs at a time.

[FourOh]

One thing I will say about the Einstein Intel GPU app - it needs about .5 CPU to run efficiently, and if you don't limit CPU projects to keep from interfering, your production will suffer greatly.

[denim]

Mine gets ~5k/day on Einstein running two WUs at a time.

I will wait a couple of days for it to average out with so many pendings, but I don't think I am going to see more than 700 credits per day combined. Either way, it is more than I was producing a couple of days ago, so every little bit counts for the team.

[FourOh]

Silly me, giving advice on Einstein's Intel GPU app and not paying attention to my results... since updating my Intel GPU drivers in late January I've had over 1000 invalids on my i7-4790k. From the Einstein Forums (http://einstein.phys.uwm.edu/forum_thread.php?id=11151&nowrap=true#138336):

"You're currently running driver 10.18.14.4080, which is known to produce bad results on HD 4600 GPUs at Einstein.

To run with the current Einstein application, you need to revert to the 10.18.10.3621 driver, as discussed in the BRP4 Intel GPU app feedback thread (http://einstein.phys.uwm.edu/forum_thread.php?id=10214)."

[zombie67]

D'oh!