Hello everyone!

There has been some really great threads lately about GPU processing, and I have been able to find great information through them. But, I am still on unsure footing, so I thought I would just post my own thread and ask a few questions all at once. Thanks to anyone who reads this, it could be a long post.

I just got windows 8 installed on my dedicated rig, and that enabled me to get my new AMD Radeon HD 7700 graphics card up and running.

Q1) With help from Juan's posting, I have this information inside of my app_config.xml file:

<app_config>
<app>
<name>setiathome_v7</name>
<gpu_versions>
<gpu_usage>0.33</gpu_usage>
<cpu_usage>0.06</cpu_usage>
</gpu_versions>
</app>
<app>
<name>setiathome_enhanced</name>
<gpu_versions>
<gpu_usage>0.49</gpu_usage>
<cpu_usage>0.06</cpu_usage>
</gpu_versions>
</app>
<app>
<name>astropulse_v6</name>
<gpu_versions>
<gpu_usage>0.51</gpu_usage>
<cpu_usage>1.00</cpu_usage>
</gpu_versions>
</app>
</app_config>

While using GPU-Z, I found that it was only until my GPU was running 3 MB tasks did it approach 98% utilization.

So my question, What will happen when my GPU decides to run an AP task, which will take up half of my card? Since the MB instructions are to use .33 for a single task, will it just run one more task and leave approximately 12% of my GPU unused during the AP task? Something feels strange here, as if I could optimize a bit better, if I only knew what I was doing :) Any hints here?

A subquestion: In my very brief testing, it seemed as if leaving three cores free helped my GPU crunch 3 MB tasks at a time. Is this true? Is it overkill? In the case of running three MB GPU tasks, how many cores should I keep free to maximize GPU usage? Besides just monitoring your GPU usage, is there any common knowledge that will point towards right or wrong?

Q2) Previously, this same rig was running with a NVIDIA GeForce 8800 GTS 512 (511MB). It did a pretty good job, with this rig netting about 200k credits in one month. (with four cores crunching as well - I don't care about credits, but its as good of a metric as any).

I have heard tales of people doing the unthinkable, and having an ATI and NVIDIA card crunching on the same rig. Some said to steer clear, while others said it was no big deal. If I can, I would LOVE to be able to have two dedicated cards crunching away, but I don't know where to begin.

I suppose first thing would be to acquire the proper NVIDIA drivers. I also suppose that this would be the easy part. However from there, from config files to xml files to free cores to everything else, I am quite clueless. Can anyone offer advice on how to get this running? How many cores would I need to keep free with two cards going? Before I just plug in the second card and risk a lot of errored WU's, I figured I'd seek some advice.

An excellent post by tbret opened my eyes a bit about CPU crunching (As did my last electric bill), and at this point I am more than ready to sacrifice CPU cores for maximum GPU optimization if that is what needs to happen.

Thanks everyone for reading, and thanks ahead of time for anyone that can share some knowledge about this topic with me!

Q1) FIFO still value, so your system will try to start the first unit received to crunch (unless some goes do high priority due th etime limit), until your GPU limit of 1 were reached, for example if you have 2 MB WU crunching a(0.66GPU) your AP if it was the next to crunch will not start (0.51GPU), if the next to crunch is a MB it will start.

subquestion: i don´t realy see a necessity to leave more than 1 core free when you crunch MB if you use an Inter CPU, on other side each AP need is own core, the supply app_config.xml allready to that with cpu_usage 1.00

If you dont crunch V6 (few V6 WU allredy left to crunch) you could eliminate the setiathome_enhanced part of the file.

Q2) Somebody please help, i have no knowledge on NV/ATI on the same host, i prefer not to mix diferent drivers on the same host.

---

Thanks very much for the info! So for the app_config information, it seems as if I can just leave it as is and no real harm will be done.

If 1 core free is all you believe is needed for MB, and the app_config.xml will free up an additional core for AP, then I guess I will put my CPU usage to 75% (I have 4), and that should leave one free.

In theory the way to make your GPU usage rise when you crunch AP is set the gpu_usage to 0.33 but if you do that you host could start to crunch up to 3 AP WU at a time, and due the way AP works that could be not a good ideia (the entire system will slow down few MB/CPU´s/GPU´s could realy handle 3 AP crunching at the same time)... so you could avoid that with max_concurrent (1 or 2) by change the AP part of the file to...

<app>
<name>astropulse_v6</name>
<max_concurrent>1</max_concurrent>
<gpu_versions>
<gpu_usage>0.33</gpu_usage>
<cpu_usage>1.00</cpu_usage>
</gpu_versions>
</app>

but you need to test to be sure it´s realy faster. Try 1 or 2 and see. In my 670 with an I5 more than 1 AP slowdown all.


If you set 75%, and you allow CPU crunching, that will startand aditional 3 CPU task when you don't crunch AP on the GPU, or 2 CPU when you crunch 1 AP on the GPU, 1 CPU task with 2 AP crunching on the GPU, etc.

---

Good idea...

If the AP WU will only use .33 of the GPU, does that mean that 2 MB's (also at .33) could be running as well if you set the max_concurrent for AP to 1?

Good idea...

If the AP WU will only use .33 of the GPU, does that mean that 2 MB's (also at .33) could be running as well if you set the max_concurrent for AP to 1?

Yes, but try, you need to test to realy see if that´s is realy faster, remember each systems is unique, there are to many diferent possible combinations of motherboards/chipsets/memory speeds/OC/GPU´s, to make a static roule.

I realy belive an 680 could easely run 1 AP + 2MB at the same time on a I5 host, 2AP+1MB don´t bet on that, 3 AP almost sure not, but i don't have a 680 to test to be sure.

But please understand, i´m not talking 3AP will not run, the GPU could run even more WU at a time (the actual limit is the memory), but due the way AP crunch they will crunch less WU/day due the overhead. Is like you could run 6 or even more MB WU at a time, but the actual production (WU/day) will be lower than 3, so I say "not run"

---

I get APs so rarely, especially ones for my GPU, that I don't think I would ever be able to do a proper test.

I may just leave my app_config the way it is, and occasionally take that 12% penalty when 1 AP and 1MB are all that the GPU can muster. This in itself would be a very rare situation, especially considering the week or more gaps between the times that AP tasks are sent out.

Thanks for the clarification, I do understand when you say "not run" meaning that the overhead cost is too high to consider it efficient.

I get APs so rarely, especially ones for my GPU, that I don't think I would ever be able to do a proper test.

I may just leave my app_config the way it is, and occasionally take that 12% penalty when 1 AP and 1MB are all that the GPU can muster. This in itself would be a very rare situation, especially considering the week or more gaps between the times that AP tasks are sent out.

I strongly sugest you to try 1AP+2MB, i´m almost sure that is the optimal configuration for the 680. An please share the results with us.

---

A 680? I may be confused, but my new card is a AMD Radeon HD 7700. Do you still recommend these settings for a AMD Radeon HD 7700?

Seeing as the 77XX series is the entry level, you should really be doing either 2MB at a time or 1&1.

I actually run that way on my GTX670.

---

A 680? I may be confused, but my new card is a AMD Radeon HD 7700. Do you still recommend these settings for a AMD Radeon HD 7700?

I can't find specs on a 7700. Did you mean 7730/7750/7770 or is it really a 7700? I don't see that model listed on newegg or amazon or at AMD's site.

<edit> I'm asking because I own a 7770 and just want to make sure I don't say something stupid. Well, anything else stupid. </edit>

A 680? I may be confused, but my new card is a AMD Radeon HD 7700. Do you still recommend these settings for a AMD Radeon HD 7700?

This is the machine that was running XP and you just upgraded to Win 8 so you could use your new 7770, correct? Looking at your results, you appear to have a driver problem. Somehow, it appears you've managed to install the newer driver but keep the AMD APP from XP...somehow...
Name: Capeverde
Vendor: Advanced Micro Devices, Inc.
Driver version: CAL 1.4.1741 (VM)
Version: OpenCL 1.1 AMD-APP (851.4)
Hopefully a fresh install of Catalyst 13.4 will solve that. Try doing a Custom Uninstall and make sure AMD APP is selected for removal. The newer Drivers have the AMD APP built in and there isn't a listing for AMD APP unless it was installed by an Older Installer, that could be the problem.

As for using your NV 8800 with your AMD 7770, there shouldn't be a problem. I have a similar setup, and up until a few days ago it was running a NV 8800 & AMD 7750. It's now running the 8800 & AMD 6770. Just make sure to install the nVidia driver first, then the AMD driver. I actually have two current Hosts running NV & AMD cards without any problems.

Seeing as the 77XX series is the entry level, you should really be doing either 2MB at a time or 1&1.

I actually run that way on my GTX670.

Thanks for your input! I'm not sure what you mean by entry level, but you are saying that the 7770 is not fit for 3 WU at a time?

Tbret: My mistake (doh!), I made several typos. It is a 7770.


TBar: Thats correct, the machine that I used to have running XP was upgraded with a brand new harddrive just so that I could run my 7770. This hard drive should have no traces of XP, the XP hard drive is unhooked.

However, from an old post I made referencing my troubles with my 7770 on XP, I followed a cataylst 12.1 download link that someone was kind enough to provide me. Do I still have a problem, even though nothing from XP should have been carried over?

Thanks for your input about the dual cards. Assuming I install the drivers in the order you've suggested, how does the rest of the installation go? Should I reinstall Lunatics, and select an option for both NVIDIA and AMD when the install dialog prompts me? Then what happens? BOINC will download WUs for both GPUs without further tweaks?

Thanks everyone for the input, it is greatly appreciated!

Your AMD APP, which is what the SETI Apps use, is Badly out of date. I'm pretty sure your times should be much better than they are, therefore, yes you have a problem. I just had a similar experience when updating from 12.8 to 13.4. The installer said the AMD APP was already installed and up to date. Being somewhat knowledgeable in such matters, I said, Whut? I did a Custom Uninstall, selected the AMD APP for removal, removed everything, then installed the 13.4 driver. That should work.

I would use the .41 installer and install only the Apps I wanted. For the 8800 you want the cuda23 version. I wouldn't run APs on the 8800, the MBs work best. I would run APs on the 7770, when available, as they work best on AMD cards.

To mherr170, when you supply something like <gpu_usage>0.33</gpu_usage>, it does not cause BOINC to enforce that limit on gpu usage for that task. It's just an estimate BOINC uses to decide what tasks to schedule. So when you say you are "wasting" 12% (should be 16%=100-33-51?) it's not really true, the gpu will run as best it can with those two tasks, could be 100% could be 60%.

Thanks TBar, I suppose I didn't quite understand the situation. I'll hunt for those drivers in the morning.

And cov_route, good to know that it isn't some strict limit I am imposing by using those settings. You can tell that a day's worth of trying to figure out how to get my rig going properly has made my head spin, as I am unable to subtract properly from 100 :)

Must head to bed, I will try to solve this driver issue then. Thanks everyone!

Thanks TBar, I suppose I didn't quite understand the situation. I'll hunt for those drivers in the morning.

And cov_route, good to know that it isn't some strict limit I am imposing by using those settings. You can tell that a day's worth of trying to figure out how to get my rig going properly has made my head spin, as I am unable to subtract properly from 100 :)

Must head to bed, I will try to solve this driver issue then. Thanks everyone!

I'm still pondering how you ended up with those driver versions. It appears you ran the Catalyst 12.1 Installer in Windows 8? Just for reference, the oldest Catalyst you should run in Win 8 is 12.8. My guess is when you ran the 12.1 Installer, if you did run it, Windows was smart enough to not install an incompatible display driver, but did install the Older AMD APP. Windows must have installed the Basic Win 8 AMD Display driver itself. Interesting...

I think the Oldest NV driver for Win 8 is around 306.97, so, make sure you install a newer version. You shouldn't have to configure the 8800. One instance of SETI version 7 will just about max it out. For ATI AstroPulse you can adjust the settings to where one instance will also max out the card. That way you don't have to worry about freeing more than one CPU core for the AstroPulse App(s). That's the way I run, one AP instance set to around 90-95% GPU load on a non-blanked AP. Works for me...

Sorry i was sleeping, my mistake i belive your GPU was a 680. So forget about 3 WU at a time on a low/mid range GPU, 2 will be better. Go for 0,51 on GPU usage for AP and and 0.49 for MB, or you could try 0.50 for each, i realy don´t like to run 2 AP at the same time on one GPU, normaly that slows down all the host.

Can't help you with NV/ATI on the same host (i only use NV) so Tbar could give you a hand on that.

---

I've been away for awhile so this may have been fixed but as I remember you have to clean install one, I believe the Nvidia first then the ATI. At the time, if you installed the ATI drivers they would play havoc with the Nvidia drivers when you tried to install them over the ATI.

I'm sure someone will correct me if I'm wrong or this has been fixed.

---

PROUD MEMBER OF Team Starfire World BOINC