The dnolan host 647520 normally runs using the integrated graphics of its motherboard. As an experiment it ran with an X1300 Pro graphics card for over a week. The comments here compare SETI results available on the Tasks for Computer web page before the X1300 test began, and from some days of running after it was put back.

Here is a graph displaying the CPU time for individual results:



Here are summary statistics by some angle range groupings others have advocated:

AR_Min	        AR_Max	        Intcpu	Xcpu	X/Int	Int#	X#
0.001000	0.050000	5498.52	5419.35	098.56%	076	089
0.120000	0.225486	5047.28	4144.05	082.10%	007	009
0.225486	0.400000	6070.92	5911.32	097.37%	159	027
0.400000	0.500000	5409.55	5459.33	100.92%	231	346
0.500000	0.750000	5104.83	4985.03	097.65%	016	004
0.750000	1.127429	4064.36	4338.92	106.76%	021	020
1.127429	10.000000	1272.74	1316.11	103.41%	094	071
						
0.001000	10.000000	4894.40	4891.02	099.93%	606	566

I think the graph and the summary statistics are best considered together. A weakness of the graphs is that average differences of several percent could easily be hidden by the general variation. A weakness of the statistics is that the samples within an angle range can be quite non-equivalent.

Happily, the .001 to .05 grouping has little if any systematic shift with AR, and is well populated. For this experiment the average CPU times matched to within 1.5%, with the add-on card case doing slightly better.

The other grouping with similar attributes is the 1.127429 to 10.0 range, for which, as it happens, the Integrated graphics variant beat the add-on card by a little more than 3%.

By contrast, the .12 to .22 range has substantial systematic shift in CPU with AR, and suffers bad mismatch between the integrated and add-on samples. The summary statistics for that range should be ignored. The same disclaimer applies, in my opinion, to the .75 to 1.127 range.

I think this experiment suggests that any advantage of the add-on card for this particular system is very small, almost certainly less than 3%, and quite possibly zero.

By the way, I think the range statistic approach can work better than it did in this case when the samples are better matched. Possibly matching can be done artificially, pausing by hand alternate samples while the first leg of the experiment is run, then releasing the matched pairs for the second leg.

An easier case is one we've seen posted in other comparisons: If two systems running simultaneously are to be compared, and both are fetching results in small batches many times a day, I think it likely that over a week or two of time the samples are likely to be largely equivalent.

This matter started from discussion the thread Building a PC for my mum and...

---

Those results are for modern PC systems.

The story may be different for PCs on which the integrated graphics system is not as advanced. On those PCs, using integrated graphics can extract a performance penalty.

See: http://www.tomshardware.com/1999/05/11/whitney/page2.html
The ASUS SiS620 board however suffers miserably in Winstone with its integrated video scoring merely 16.5 Winstones. It is apparent the integrated video on the SiS620 steals performance from the ASUS platform. I went ahead and stuck a V550 PCI board in the same platform and the Winstone score jumped up to 19.0!

I know this quite an old link, but it shows that on older motherboards, using integrated graphics instead of discreet graphics can significantly reduce the PC's performance in office applications (and probably in SETI@Home too).

... it shows that on older motherboards, using integrated graphics instead of discreet graphics can significantly reduce the PC's performance in office applications (and probably in SETI@Home too).

I can't see how Winstone results (designed to represent typical workloads of current apps back when they were developed) can be extrapolated to crunching. Machine requirements for e.g. MS Office and for SETI performance are quite different as far as I can see.

F.

---

... it shows that on older motherboards, using integrated graphics instead of discreet graphics can significantly reduce the PC's performance in office applications (and probably in SETI@Home too).

I can't see how Winstone results (designed to represent typical workloads of current apps back when they were developed) can be extrapolated to crunching. Machine requirements for e.g. MS Office and for SETI performance are quite different as far as I can see.

F.

On that ASUS motherboard, the integrated graphics ate up a fairly significant part of the CPU's processing power. I think that would mean some CPU procressing power is made unavailable to SETI (not just office applications).

What has not been mentioned in this thread is that the machine in question is set to turn off graphics after 10 to 20 minutes, per this post in the "Building a PC for my mum and..." thread. If the graphics are truly disabled, it makes no difference whether a video card is installed or not.

What wasn't stated was how often per day the graphics are turned back on, but as it's a dedicated 24/7 crunching machine I'd guess seldom.

                                                                Joe

What has not been mentioned in this thread is that the machine in question is set to turn off graphics after 10 to 20 minutes, per this post in the "Building a PC for my mum and..." thread. If the graphics are truly disabled, it makes no difference whether a video card is installed or not.

What wasn't stated was how often per day the graphics are turned back on, but as it's a dedicated 24/7 crunching machine I'd guess seldom.

                                                                Joe

Good point. Some folks have been advocating that new-build dedicated crunchers should include a graphics card in preference to integrated graphics for performance reasons. I think this set of observations casts some doubt on whether that is a cost/performance effective choice on new-build machines.

The case of another machine making extensive use of its graphics for non-SETI purposes could well be quite different.

---

What wasn't stated was how often per day the graphics are turned back on, but as it's a dedicated 24/7 crunching machine I'd guess seldom.

                                                                Joe

It frequently goes days without having the graphics turned on...

FYI

-Dave

---

I have two machines running BOINC with S@H

Macintosh G4; 867Mhz PPC with 2MB L3
1GB SDRAM, Unlimited disk space ATA133 @ 7200RPM
Video: ATI 9800 PRO
OSX Tiger

SONY VIAO; 1.73 CentrinoCore2Duo (T5300) with 2MB L2
1GB PC5300, 120GB SATA @ 5300RPM
Video: Intel 950 Graphics
Windows VISTA


Logically the Dual Core should be faster with BOINC, but with Graphics enabled, it sucks. Memory is shared and the whole thing lags ever time it refreshes one of those 3D frames even with the settings Gimped...
The Old G4 has so far cruised right passed the VAIO, even with Graphics!
In my opinion, its a No brainer, a Good Graphics Card does make a difference.
If yours is built in, just make sure its not a stock Intel based accelerator! After that all should be Good.


Edit:

Know what? I retract that. I must have debugged the settings without knowing it. SETI is now blazing away like Id expect on a Dual Core, and its only using about 65% on the spedometer. I'm sure disabling allot of the 3d junk helped. So at least we know that the card is REALLY only important if you intend to Display while SETI works. I know, I enjoyit.

If yours is built in, just make sure its not a stock Intel based accelerator!

FYI - the on board graphics on my system (used for the test in the original message) is Intel (G31/33)

-Dave

---

My system memory speed:

With external video




with onboard video



As you see difference in memory read about 2%.

---