- Dec 31, 2007
- 6,223
- 1
- 25,780
Right now my company is evaluating a single board computer to run in a kiosk-type system with a touchscreen LCD. Due to some old fogeys in the hardware ordering process, our first evaluation unit was a P4 3GHz on an i865 mobo paired with 1GB of <i>single-channel DDR266</i>. Ouch. Try as I might to explain that such a processor is best paired with DC DDR400, I was only partially successful. So what was the second eval setup? 1GB dual-channel DDR266. :O It figures, no?
As funny (or not) as this is, I was able to OC the RAM to DDR400 with only slightly worse timings. (BIOS fought me on this by denying <i>any</i> manual settings to the timings and putting them to the absolute worst in manual mode, but I got close to the same timings in the end by leaving the timings taken from SPD and just adjusting the clock speed from 266 to 400.) So I was able to get some fairly interesting benchmarks run.
Why is this interesting? Because now you get to see just exactly how badly things perform when you starve a P4 by pairing it with less memory bandwidth than it needs.
Setup 1 = SC DDR 266 CAS2.5 6-3-3 (1065 max MB/s in Sandra)
Setup 2 = SC DDR 400 CAS2.5 7-4-4 (1888 max MB/s in Sandra)
Setup 3 = DC DDR 266 CAS2.5 6-3-3 (2668 max MB/s in Sandra)
Setup 4 = DC DDR 400 CAS2.5 7-4-4 (3979 max MB/s in Sandra)
SiSoft Sandra CPU ALU:
1) 7270 MIPS (5.96% perf loss)
2) 7446 MIPS (3.69% perf loss)
3) 7710 MIPS (0.27% perf loss)
4) 7731 MIPS (0.00% perf loss)
Prime95 total benchmark runtime:
1) 508.135ms (16.29% perf loss)
2) 440.684ms (03.48% perf loss)
3) 430.892ms (01.29% perf loss)
4) 425.333ms (00.00% perf loss)
PCMark2002 CPU:
1) 6075 (2.73% perf loss)
2) 6186 (0.67% perf loss)
3) 6213 (0.22% perf loss)
4) 6227 (0.00% perf loss)
PCMark2002 Memory:
1) 4828 (44.17% perf loss)
2) 6204 (28.25% perf loss)
3) 7057 (18.39% perf loss)
4) 8647 (00.00% perf loss)
3DMark2001:
1) 2078 (31.87% perf loss)
2) 2545 (16.56% perf loss)
3) 2694 (11.67% perf loss)
4) 3050 (00.00% perf loss)
I know that the benchmarks used aren't the most stunning, but they're what I had on hand. (Well, the graphics benchmark was chosen because the onboard graphics are DX7 so the score isn't skewed by lackings in the hardware support.) I'd loved to have done some game testing for more real-world results, but I have a hard enough time not getting into trouble while running 3DMark. Installing and running game evals at work could really get me into some hot water, no matter how useful it might be for benchmarking systems.
What is interesting to me is:
1) In some cases the actual performance loss is fairly small, but in others is huge. For example, SiSoft Sandra FPU was completely unaffected. Yet SiSoft Sandra ALU seemed more affected by SC vs DC than actual bandwidth. And then Prime95 and PCMark2002-CPU didn't seem all too bad off until they hit rock bottom memory performance.
2) The performance loss scales fairly normally until you hit SC DDR266, where it kills the performance badly. So it seems that the P4 not only performs worse when starved for memory bandwidth, but it also has a minimum line where it <i>really</i> keels over dead without enough bandwidth. (Because of this I wish that I'd had the time to also test with DDR333 settings.)
I just thought that it was an interesting little experiment. I mean we all knew that pairing a P4 with lesser memory would hinder performance, but until now I never really knew how much so in some cases, and how little in others.
I also have to wonder how much the onboard GPU skewed the 3D results. :\ Too bad the single board PC only has PCI and ISA slots.
<pre><font color=purple><i>Jesters do oft prove prophets.</i> -Regan in
King Lear (Act V, Scene iii) by William Shakespear</font color=purple></pre><p>@ 187K -> 200,000 miles or bust!
As funny (or not) as this is, I was able to OC the RAM to DDR400 with only slightly worse timings. (BIOS fought me on this by denying <i>any</i> manual settings to the timings and putting them to the absolute worst in manual mode, but I got close to the same timings in the end by leaving the timings taken from SPD and just adjusting the clock speed from 266 to 400.) So I was able to get some fairly interesting benchmarks run.
Why is this interesting? Because now you get to see just exactly how badly things perform when you starve a P4 by pairing it with less memory bandwidth than it needs.
Setup 1 = SC DDR 266 CAS2.5 6-3-3 (1065 max MB/s in Sandra)
Setup 2 = SC DDR 400 CAS2.5 7-4-4 (1888 max MB/s in Sandra)
Setup 3 = DC DDR 266 CAS2.5 6-3-3 (2668 max MB/s in Sandra)
Setup 4 = DC DDR 400 CAS2.5 7-4-4 (3979 max MB/s in Sandra)
SiSoft Sandra CPU ALU:
1) 7270 MIPS (5.96% perf loss)
2) 7446 MIPS (3.69% perf loss)
3) 7710 MIPS (0.27% perf loss)
4) 7731 MIPS (0.00% perf loss)
Prime95 total benchmark runtime:
1) 508.135ms (16.29% perf loss)
2) 440.684ms (03.48% perf loss)
3) 430.892ms (01.29% perf loss)
4) 425.333ms (00.00% perf loss)
PCMark2002 CPU:
1) 6075 (2.73% perf loss)
2) 6186 (0.67% perf loss)
3) 6213 (0.22% perf loss)
4) 6227 (0.00% perf loss)
PCMark2002 Memory:
1) 4828 (44.17% perf loss)
2) 6204 (28.25% perf loss)
3) 7057 (18.39% perf loss)
4) 8647 (00.00% perf loss)
3DMark2001:
1) 2078 (31.87% perf loss)
2) 2545 (16.56% perf loss)
3) 2694 (11.67% perf loss)
4) 3050 (00.00% perf loss)
I know that the benchmarks used aren't the most stunning, but they're what I had on hand. (Well, the graphics benchmark was chosen because the onboard graphics are DX7 so the score isn't skewed by lackings in the hardware support.) I'd loved to have done some game testing for more real-world results, but I have a hard enough time not getting into trouble while running 3DMark. Installing and running game evals at work could really get me into some hot water, no matter how useful it might be for benchmarking systems.
What is interesting to me is:
1) In some cases the actual performance loss is fairly small, but in others is huge. For example, SiSoft Sandra FPU was completely unaffected. Yet SiSoft Sandra ALU seemed more affected by SC vs DC than actual bandwidth. And then Prime95 and PCMark2002-CPU didn't seem all too bad off until they hit rock bottom memory performance.
2) The performance loss scales fairly normally until you hit SC DDR266, where it kills the performance badly. So it seems that the P4 not only performs worse when starved for memory bandwidth, but it also has a minimum line where it <i>really</i> keels over dead without enough bandwidth. (Because of this I wish that I'd had the time to also test with DDR333 settings.)
I just thought that it was an interesting little experiment. I mean we all knew that pairing a P4 with lesser memory would hinder performance, but until now I never really knew how much so in some cases, and how little in others.
I also have to wonder how much the onboard GPU skewed the 3D results. :\ Too bad the single board PC only has PCI and ISA slots.
<pre><font color=purple><i>Jesters do oft prove prophets.</i> -Regan in
King Lear (Act V, Scene iii) by William Shakespear</font color=purple></pre><p>@ 187K -> 200,000 miles or bust!
Facebook
Twitter
Instagram