I am running stock seti (at least for the time being).

I just got a Radeon HD 5670 and have on order a Gtx 750 Ti. The Radeon was unexpected but since I have two PCexpress 16 slots on my Xeon motherboard I figured "why not". I currently have a Nvidian Quadro FX 1800 in the slot the Gtx will go in.

I have been studying the CUDAxx .cfg documentation as well as the Ati/Radeon documentation. And I think I now have a handle on what needs to be set where.

The gotcha is the Ati/Radeon documentation heavily recommends running at least one of my cpu cores exclusively for the card. If you have two Radeon cards it recommends at least 2 cpus. The Nvidia documentation is less clear about if you should/shouldn't dedicate a cpu to a gpu.

Another thread in the community is quite specific about Astropulse. It recommends running a full cpu in support of a gpu based Astropulse wu. Yes, I have been getting basic Astropulse work units that don't use a gpu.

So it is pretty clear. My baseline needs to be no more than 1 task per gpu. I am currently experimenting with 0.24 for the cpu on setiathome_v7 in the app_config.xml file because that allows me to process 8 basic seti units/basic astropulse units while processing 2 more gpu units. The question is how much will is slow down the gpu based processing?

If you have your favorite setups for app_config.xml, mbcuda-7.00-cudaXX.cfg and mb_cmdline-7.03-open_ati?_sah.txt (the ? means an optional 5 goes there) you would like to post please be my guest.

Thanks,
Tom

---

A proud member of the OFA (Old Farts Association).

The gotcha is the Ati/Radeon documentation heavily recommends running at least one of my cpu cores exclusively for the card. If you have two Radeon cards it recommends at least 2 cpus. The Nvidia documentation is less clear about if you should/shouldn't dedicate a cpu to a gpu.

To keep the gpu busy, a cpu needs to be available to handle its data needs quickly/timely. Not a dedicated cpu, only a mix of work that ensures that most of the time a cpu is available quickly. Since the gpu can perform many more calculations than a cpu per unit time, it is better to waste some cpu to keep the gpu busy.

Another thread in the community is quite specific about Astropulse. It recommends running a full cpu in support of a gpu based Astropulse wu. Yes, I have been getting basic Astropulse work units that don't use a gpu.

I have found this to be true on my fastest machine. Even a dedicated cpu isn't fast enough to keep even the modest GTX750Ti gpu completely busy.

So it is pretty clear. My baseline needs to be no more than 1 task per gpu. I am currently experimenting with 0.24 for the cpu on setiathome_v7 in the app_config.xml file because that allows me to process 8 basic seti units/basic astropulse units while processing 2 more gpu units. The question is how much will is slow down the gpu based processing?

I don't have any AMD processors on my systems, only Intel, but here is what I have found. Based on your cpu and gpu configuration, your mileage may vary.

On my main cruncher (http://setiathome.berkeley.edu/show_host_detail.php?hostid=7265951) with four GTX750Ti cards, Astropulse will use all of a cpu (and want more!), and about 80% +/- of a gpu. So when enough Astropulse and MB units are available, I run one of each in each gpu, and leave 5 cpus to do the servicing. The remaining 3 cpus I leave for another project. With this mix, I get 95%+ utilization of the gpus, and 98% of all the cpus. If only MB are available, two run on each gpu giving 95% usage, and use only 1 cpu to service them, leaving the other 7 cpus for another project, and with cpu 95-98% busy. Thus running 8 MB need LESS than one cpu. The way I got this mix was trial and tweak until the % usage were both high. Here are the parameters:

<app_config>
<app>
<name>setiathome_v7</name>
<gpu_versions>
<gpu_usage>0.49</gpu_usage>
<cpu_usage>0.2</cpu_usage>
</gpu_versions>
</app>
<app>
<name>astropulse_v6</name>
<gpu_versions>
<gpu_usage>0.51</gpu_usage>
<cpu_usage>1.1</cpu_usage>
</gpu_versions>
</app>
</app_config>

I think you will have to "play" with the numbers a bit to find the best ones for your configuration and projects you support. If you "starve" your gpus for cpu time, your performance will be less than optimal, so trial and tweak is the name of this game.

[Edit to clarify number of cpus used with MB]

I just had to give the Radeon HD 5670 back :( the family seller had next day regrets about losing his backup video card.

So I still have GTX 750 Ti coming which should replace my Nvidia ?? 1800 with only some simple changes in the *.cfg files.

But in another thread where I was struggling to get a tiny APU system with a small radeon card embedded, that it turned out I didn't have a high enough version of the drivers, once I got that fixed he told me to:

"...The next step is you're going to have to delete the apps compilations, they were done on the incompatible driver,

These are a number of files in your project directory ending in .bin , .bin_v7 and .wisdom, suspend Boinc usage, delete them then resume Boinc usage, they'll be regenerated by the new APP runtime...."

I just ran across an elderly technology with a TeraFlop ATI video card (FireStream they call it) for $34.99 so I ordered it. I suspect that if/once I install it on the Xeon it will be in the same place as my other machine. You have to have the code that runs on the Radeon video cards generated with that specific card/drivers in mind.

Right?

Thanks,
Tom

---

A proud member of the OFA (Old Farts Association).