Two ATI Radeon HD 5770 in a Mac Pro 3,1?

Tigers_Dave
Tigers_Dave
Joined: 25 Mar 09
Posts: 21
Credit: 698,209,778
RAC: 3,228,075
Topic 196536

Hi All:

One of my computers (ID: 3217417 ) is an "Early 2008" Mac Pro 3,1 (2 x Quad-core 2.8 GHz Xeon E5462) with 10 GB RAM. I am using OS X 10.8.2 and BOINC 7.0.31 (x86). The original video card (ATI Radeon HD 2600 XT) began to fail and I replaced it with an ATI Radeon HD 5770 (1024 MB). Following the installation of the new video card, the overall RAC for this unit (not simply the E@H RAC) substantially increased, probably due to the fact that the new video card allows me to process OpenCL E@H work units. This model of Mac Pro can be equipped with a second ATI Radeon HD 5770 card. Would E@H recognize a second card and thereby enable me to experience an additional increase in E@H processing capability? Any downside? TIA!

Gary Roberts
Gary Roberts
Moderator
Joined: 9 Feb 05
Posts: 5,491
Credit: 63,171,103,073
RAC: 54,261,141

Two ATI Radeon HD 5770 in a Mac Pro 3,1?

Quote:
... Would E@H recognize a second card and thereby enable me to experience an additional increase in E@H processing capability?


Projects can only know about hardware if the hardware is recognised by BOINC. That depends on BOINC version and there should be no problem with your up-to-date version. So, yes, a second graphics card should properly be recognised. Unless it was an OS X requirement, you should even be able to use a different model card if you wanted to.

I had a quick look at the tasks list for that machine and I noticed quite a variation in run times - essentially between 4Ksecs and 10Ksecs with all shades in between. Are you using the 'GPU utilization factor of BRP apps' preference setting to run more than one task simultaneously? I suspect not but I thought I'd ask. I'm wondering if the variability in run time is an indication that the GPU isn't being fed as efficiently as it could be. Your machine has 8 CPU cores. It would be really worthwhile to set the computing preference 'On multiprocessors, use at most ...' to 87.5% (just for that host - perhaps use a different venue so as not to affect other hosts) so that you had BOINC using only 7 cores and one core available to feed the GPU. That can make a substantial difference to GPU performance. The increase in GPU productivity can easily outweigh the loss of a CPU core.

If you do go ahead with the second GPU, you would certainly need to free up a CPU core - probably even two. You'd need to experiment with both the number of free cores and the GPU utilisation factor to get the most productive combination.

Quote:
Any downside?


Probably only extra heat and increased power consumption :-).

Cheers,
Gary.

Tigers_Dave
Tigers_Dave
Joined: 25 Mar 09
Posts: 21
Credit: 698,209,778
RAC: 3,228,075

Thanks for the advice, Gary.

Thanks for the advice, Gary. Will tweak 'GPU utilization factor of BRP apps' preference as you suggested. Will see how that turns out before I add another ATI HD 5770 to that computer. I am also going to consider putting an ATI HD 5770 in an older MacPro1,1 - the (original) video card in that box has been a little hinky.

Also, thanks for your comments in one of the BOINC message boards (posts 42178 and 42185, Jan 22/23, 2012) regarding the ATI Radeon HD 4670 in an iMac. It helps explain the problems I have been having with Collatz@Home with that video card in an i3 iMac.

Gamboleer
Gamboleer
Joined: 5 Dec 10
Posts: 173
Credit: 168,389,195
RAC: 0

Try setting the "GPU

Try setting the "GPU utilization factor of BRP apps" to 0.5 (this is done by editing your Einstein@Home preferences on your account on the website) and the % of CPUs used to 87.5% on that 3,1; this will run two GPU tasks on the 5770 with the one free core.

On my 1,1 with a 5770 I use 75% of CPUs with the 0.5 GPU utilization, since it's only a 4-core model, and it works well, doing 2 simultaneous Arecibo tasks in just under 2 hours each. Trying to run more than 2 GPU tasks made the card choke.

I haven't seen the motherboard on a 3,1 but remember the 5770s are double-wide cards, and on the 1,1 the only 16x slot is slot 1, at the bottom; if you're running 16x in Slot 1, Slot 4 (which can't hold a 5770 because it's too close to the hard drives) is the only one that will run at 8x. If the 3,1 motherboard is the same layout, I think your only possible configuration is a card in Slot 1 and a second in Slot 3, and I think if you do that you're at 4X in Slot 3. This may impact the card's performance.

My suggestion: pull the 5770 and try it in your 1,1 to make sure it works OK (they seem to work on most 1,1 systems, but require 10.6.4 or better to function as 3D cards and not just 2D). If it does, buy a 5870 and put that in the 3,1; it's about 50% faster than the 5770 and you'll get full 16x PCI-E performance from it since it will be in Slot 1. That may be better than running two 5770s, when the second card will have to reside in a 4x slot.

Gary Roberts
Gary Roberts
Moderator
Joined: 9 Feb 05
Posts: 5,491
Credit: 63,171,103,073
RAC: 54,261,141

RE: .. Will tweak 'GPU

Quote:
.. Will tweak 'GPU utilization factor of BRP apps' preference as you suggested.


Make sure you free up at least one CPU core before you do that. In fact you should free up one core first and wait to see how much benefit that gives on its own. Then you should try running two simultaneous GPU tasks (still with one free core) to see if there is further improvement. Then you should free up a second core to see if the further loss of a CPU can still be outweighed by an increase in GPU output. At least you should then know what is likely to be best.

I suspect (admittedly from very scant direct experience) that you may find that two GPU tasks and one free CPU core gives you the best output but it could be that 2 free CPU cores does even better. You need to experiment systematically to find out for sure.

Cheers,
Gary.

Tigers_Dave
Tigers_Dave
Joined: 25 Mar 09
Posts: 21
Credit: 698,209,778
RAC: 3,228,075

Gary and Alec: Thanks for

Gary and Alec:

Thanks for the additional advice. I have changed the 'GPU utilization factor of BRP apps' preference to 0.5 and the '% of CPUs used' preference to 87.5%. Both of these preference changes were applied to the 'work' venue and this venue is used only by the MacPro3,1. I updated E@H and, as expected, I am now crunching 7 CPU tasks. However, I am crunching only 1 E@H GPU task. I expected that altering the 'GPU utilization factor of BRP apps' preference to 0.5 would allow me to crunch 2 E@H GPU tasks. Am I mistaken, or does it take a while for this preference change to have an effect? TIA!

Gary Roberts
Gary Roberts
Moderator
Joined: 9 Feb 05
Posts: 5,491
Credit: 63,171,103,073
RAC: 54,261,141

RE: ,.. I expected that

Quote:
,.. I expected that altering the 'GPU utilization factor of BRP apps' preference to 0.5 would allow me to crunch 2 E@H GPU tasks. Am I mistaken, or does it take a while ...


The change applies to new tasks that are sent to you. I'm not actually sure if old tasks already in your cache have to be cleared first or whether you just have to download new work for the change to apply. I thought you had to finish the cached tasks as singles and then the new tasks would be done as doubles but I did see a recent comment that old tasks could be done as doubles once the setting change had been communicated to the client through the arrival of new work. I don't know whether that is correct or not.

Are your old tasks crunching any faster (as singles) with the freeing of a CPU core?

Cheers,
Gary.

archae86
archae86
Joined: 6 Dec 05
Posts: 3,008
Credit: 4,844,565,370
RAC: 3,363,344

RE: I did see a recent

Quote:
I did see a recent comment that old tasks could be done as doubles once the setting change had been communicated to the client through the arrival of new work. I don't know whether that is correct or not.

I have both read such comments and actually observed such behavior on my host, I believe. So it can take quite a while for the change to be effective, or it can be quite rapid.

Tigers_Dave
Tigers_Dave
Joined: 25 Mar 09
Posts: 21
Credit: 698,209,778
RAC: 3,228,075

Thanks, Gary and archae86,

Thanks, Gary and archae86, for the illuminating comments. The run-time for these BRP tasks is in the range of 3.8K-4.2K seconds, which is a significant improvement. At this point, because of the fact that this machine does work for four different BOINC projects (and has been doing a fair amount of "real" work), it is difficult to say at this point whether the improvement in processing the BRP tasks is greater than the loss of productivity caused by dedicating one of my CPU cores. For the past month, this machine has averaged 12K BOINC credits/day. I'll continue to monitor this machine's productivity and will post a followup.

Thanks for advising me that I may need to clear the cache of BRP tasks present at the time of changing the E@H preferences before I experience the full effects of the changing those preferences. My back-of-the envelope calculations (90 tasks @ 4K seconds run-time/task) indicate that it will take approximately 4 more days to clear the cache. Will post an update after that cache is cleared.

Finally, thanks to Alec for suggesting that I buy an ATI Radeon HD 5870. However, $500 is bit steep for what essentially is an upgrade to my MacPro 1,1. I'll try to limp along and save my pennies until I can afford a CTO i7 iMac ($1800) or the long-awaited major upgrade to the Mac Pro that Tim Cook (War Eagle!) promised us for 2013.

Gamboleer
Gamboleer
Joined: 5 Dec 10
Posts: 173
Credit: 168,389,195
RAC: 0

You're welcome. Not that I

You're welcome. Not that I think this will change your mind, but BH Photo has the 5870 for $440:

http://www.bhphotovideo.com/c/product/726539-REG/Apple_MC743ZM_A_ATI_RADEON_HD_5870.html

I bought my 5770 from them for about $225:

http://www.bhphotovideo.com/c/product/726537-REG/Apple_MC742ZM_A_ATI_Radeon_HD_5770.html

Good luck, whatever you decide!

Tigers_Dave
Tigers_Dave
Joined: 25 Mar 09
Posts: 21
Credit: 698,209,778
RAC: 3,228,075

My MacPro3,1 has cleared its

My MacPro3,1 has cleared its cache of older E@H GPU tasks. Alas, I am still only processing one GPU task at a time. However, the decreased processing time for these tasks appears to balanced out the loss of one of my CPU cores. Thus, the overall BOINC throughput on this machine has increased from roughly 12,500 units/day to 13,500 units/day. This is not very impressive compared to what the uber-crunchers are getting from the latest and greatest boxes. However, it is significantly better than the 6,000 units/day that I was getting before replacing my stock ATI Radeon HD 2600 XT with the ATI Radeon HD 5770. Thanks again to everyone for the helpful advice!

Comment viewing options

Select your preferred way to display the comments and click "Save settings" to activate your changes.