Overclocking and Benchmarking: CPU, DDR or both?

[Jim]

Recently there's been much discussion (read arguement) about what really counts when configuring/benching a system. Such as P4"B" and "C" chips, DDR MHZ, DDR timings, CPU MHZ, "synthetic benchmarks" or "real world benchmarks" when looking for optimum system performance. I thought I'd run a suite of benchmarks using one identical system and try and explore as many of these variables as possible. Since much of this recent discussion centers on DDR speeds and timings and the effect of high DDR MHZ and latencies with 1:1 and 5:4 ratios that will be my focus.

[Jim]

In order to test as many of these variables as possible with one system so all benchamrks would be comparable the key component in testing was the DDR. Obviously DDR that would do 2-2-2-5 timings at both 400MHZ and 500MHZ speeds for comparison would be optimal. However, as far as I know this doesn't exist for the average user with typical voltages or hardware. These tests are all about reasonable benchmarks applicable to someone building a system today that might be orderred from your favorite online retailer. In order to achieve this and examine system performance at both 400MHZ DDR and 500MHZ DDR speeds certain compromises were necassary. As already stated DDR that will do 2-2-2-5 at any speed would make for an easy comparison but this is not possible. Memory that would perform well with good timings at both spectrums was necassary to make these comparisons viable.

Enter OCZ 3700 Gold.

http://users.adelphia.net/~jimscomp/3700_Gold.JPG

[Jim]

I found that the OCZ memory was capable of performing at all the speeds/settings I wanted to compare at a satisfactory bios setting of 2.5-3-3-7. Not 2-2-2-5 but not 3-4-4-8 either. ;) Note also this is 2X256MB memory, not 2X512MB.


First, we'll take a look at the system used.

Asus P4C800-E Deluxe 1010 bios
P4C 2.4C/Swiftech MCX4000 w/92mm Sunon
OCZ 2X256 3700 Gold DDR
eVGA FX5900 non Ultra
WD Raptor 36gb/Sata
Enermax 431 watt PS
Toshiba 16X DVD/Plextor 40X CDRW
XP Pro/SP1
45.23 Detonators/Directx 9.0b
HT enabled, PAT enabled, Turbo DDR setting in bios, Quality setting for FX5900, FX5900 @ default clock of 400/850. CPU vcore auto, DDR vcore 2.85v, for all tests.
ALL components were purchased through retail channels by myself and are not "hand picked".

[Jim]

In order to compare performance with both a 1:1 ratio as well as a 5:4 ratio which is often used to overclock a P4C chip I used both a 201FSB setting as well as 251 FSB settings for my initial tests. 201FSB (1:1 ratio = 402 DDR MHZ) can be compared to 251FSB (5:4 ratio = 401.6 DDR MHZ) . These settings as the DDR would not complete all tests at 200FSB. 3 speed/setting combinations were used to further compare performance. 201FSB 1:1 DDR @ 402MHZ, 251FSB with a 5:4 divider DDR @ 401.6MHZ and 251FSB 1:1 DDR at 502MHZ. With the DDR running at the same speed timings of 2.5-3-3-7 and the CPU at 201FSB and @ 25F1SB as well as both the CPU and DDR overclocked we can examine where the performance gains lie when overclocking.
Since most PC4000 memory available today is rated at 3-4-4-8 timings I also looked at 250FSB and 1:1 settings for the CPU and DDR to compare the difference in system performance using both 2.5-3-3-7 and 3-4-4-8 DDR timings.
Unfortunately with this system a 2.4B chip comparison is not possible using these parameters.
With these five tests hopefully one can conclude the importance of CPU and DDR speeds, as well as DDR timings, especially when benchmarking to demonstrate DDR performance.
Also, the validity of synthetic versus real world benchmarks will be looked at.

[Jim]

Seven benchmarks were used for this comparison. Four synthetic and three real world.


Synthetics:


http://www.bleedinedge.com/images/thugs/ico-a32.gif Aida32 v3.70 read/write


http://www.bleedinedge.com/images/thugs/ico-pcm2k2.gif PCMark2002 memory


http://www.bleedinedge.com/images/thugs/ico-sss.gif Sandra v7973 unbuffered


http://www.bleedinedge.com/images/thugs/ico-3dm2k1.gif 3DMark2001
Default settings

Real world:


http://www.bleedinedge.com/images/thugs/ico-ut2k3.gif UT2003 Demo 1280x960, Fly / Bot
1280X960 Resolution


http://www.bleedinedge.com/images/thugs/ico-x2.gif X2 Rolling Demo
Default settings


SUPER_PI
PI calculated to 2M

[Jim]

First benchmarks up 201,251 FSB 1:1 and 251 FSB 5:4 ratio @ 2.5-3-3-7

201 FSB 1:1 = CPU 2412 MHZ/ DDR 402 MHZ
251 FSB 5:4 = CPU 3012 MHZ/ DDR 401.6 MHZ
251 FSB 1:1 = CPU 3012 MHZ/ DDR 502 MHZ


Synthetics

AIDA Read/Write
201 FSB 1:1 = 4880/1750
251 FSB 5:4 = 5089/1750
251 FSB 1:1 = 5773/1992

PCMark
201 FSB 1:1 = 8568
251 FSB 5:4 = 9848
251 FSB 1:1 = 10600


SANDRA
201 FSB 1:1 = 2547/2656
251 FSB 5:4 = 2607/2684
251 FSB 1:1 = 3225/3348

3DMark 2001
201 FSB 1:1 = 14033
251 FSB 5:4 = 15877
251 FSB 1:1 = 16228


Real World

UT2003 fly/bot
201 FSB 1:1 = 172/62
251 FSB 5:4 = 190/75
251 FSB 1:1 = 193/78

X2 Rolling Demo
201 FSB 1:1 = 79.38
251 FSB 5:4 = 93.56
251 FSB 1:1 = 95.37

Super_Pi
201 FSB 1:1 = 2:13
251 FSB 5:4 = 1:50
251 FSB 1:1 = 1:45

It's obvious here that for system performance CPU MHZ rules. Aida, PCMark and Sandra register large performance increases when moving to 251FSB 1:1 ratio but the real world benchmarks have already shown the performance gained with the CPU set to 251 FSB and the DDR at 5:4.

[Jim]

Next up a 250 FSB comparison with a 1:1 ratio and 2.5-3-3-7 versus 3-4-4-8 timings.
250FSB 1:1 = CPU 3000 MHZ/ DDR 500 MHZ at 2.5-3-3-7 and 3-4-4-8 comparison

Synthetics

AIDA Read/Write
2.5-3-3-7 = 5726/1996
3-4-4-8 = 5596/2156

PCMark
2.5-3-3-7 = 10669
3-4-4-8 = 10574


SANDRA
2.5-3-3-7 = 3324/3368
3-4-4-8 = 3171/3166

3DMark 2001
2.5-3-3-7 = 16279
3-4-4-8 = DNF

Real World

UT2003 fly/bot
2.5-3-3-7 = 193/78
3-4-4-8 = 192/76

X2 Rolling Demo
2.5-3-3-7 = 94.48
3-4-4-8 = 93.90

Super_Pi
2.5-3-3-7 = 1:46
3-4-4-8 = 1:48

Here we see timings have a minimal effect on real world performance but the synthetic benchmarks do a more accurate job of indicating the effect of DDR timing tweaks. With a small peformance hit realized real world from looser DDR timings the possible ability to overclock the CPU higher becomes very attractive with the looser timings.

[Jim]

Conclusion


For real world performance, CPU MHZ rules (Duh). In our first set of tests we saw the greatest performance gains immediately once the FSB and CPU MHZ were raised, even with a 5/4 DDR ratio being used. Synthetic benchmarks for DDR failed to show accurately these gains simply because they're better at showing DDR performance improvements that don't translate to real world benchmarks as dramatically as CPU MHZ does. I think when reviewing and overclocking DDR this can be misleading as both DDR and CPU performance gains are sometimes shown together whereas the CPU is responsible for the lions share of these performance gains.

In our second set of tests the synthetics did a better job of indicating what tweaks might give better system performance when fine tuning a system. Real world benchmarks showed little peformance difference when comparing DDR timings making it more difficult to judge the effect of changes made. Since identical CPU and DDR chips will perform differently based mainly what batch they're from or luck of the draw, the only way to test and fine tune your system and to find its potential is to use a balanced set of benchmarks and tweak your system until the best possible results are realized.
Happy Overclocking! PF :D

[Jim]

Overtime

What we haven't looked at yet is the ability of the P4C chip to overclock far beyond the OCZ3700's ability to keep up using a 1:1 ratio. Lets see how far we can go with a 5:4 ratio past the benchmarks already shown and see if we can leave 1:1 in the dust.
For comparison I'll relist the 251 FSB 1:1 2.5-3-3-7 tests results. I'll also see if I can sqeeze out a little more FSB overclock with a 1:1 DDR ratio. To achieve this DDR timings were set to 2.5-4-4-8.

282FSB 5:4 = CPU 3384 MHZ/ DDR 451 MHZ @ 2.5-3-3-7
251FSB 1:1 = CPU 3012 MHZ/ DDR 502 MHZ @ 2.5-3-3-7
254FSB 1:1 = CPU 3048 MHZ/ DDR 508 MHZ @ 2.5-4-4-8

Synthetics

AIDA Read/Write
282 FSB 5:4 = 5712/1917
251 FSB 1:1 = 5773/1992
254 FSB 1:1 = 5769/1931

PCMark
282 FSB 5:4 = 11063
251 FSB 1:1 = 10600
254 FSB 1:1 = 10583

SANDRA
282 FSB 5:4 = 2952/3009
251 FSB 1:1 = 3225/3348
254 FSB 1:1 = 3057/3169

3DMark 2001
282 FSB 5:4 = 17003
251 FSB 1:1 = 15877
254 FSB 1:1 = 16250

Real World

UT2003 fly/bot
282 FSB 5:4 = 198/85
251 FSB 1:1 = 193/78
254 FSB 1:1 = 193/78

X2 Rolling Demo
282 FSB 5:4 = 100.88
251 FSB 1:1 = 95.37
254 FSB 1:1 = 95.48

Super_Pi
282 FSB 5:4 = 1:38
251 FSB 1:1 = 1:45
254 FSB 1:1 = 1:46

[Jim]

Final thoughts

Once again CPU MHZ rules the day. Interesting also was what a great job Aida, PCMark and Sandra did showing the loss of DDR bandwidth when the speeds were set to 254 FSB 1:1 @ 2.5-4-4-8. This is the strongest arguement I can make for the validity of Synthetic benchmarks. The few extra CPU cyles were enough to keep the real world and 3DMark results equal to or just above 251 FSB 1:1 @ 2.5-3-3-7.
The only other comparisons versus the P4's ability to overclock using a 5:4 ratio would be a comparison of 2X256 high quality PC4000 DDR sticks ran well over the 508 MHZ 1:1 ratio the OCZ3700 would do. Stay tuned if I sneak a pair past the doorstep and I'll update the thread.

PF :D

[eva2000]

very nice indeed :)

Just how many users can get their 2.4C stable at reasonable voltages on air @284fsb 5:4 versus @250-260fsb 1:1

btw have you tried comparing with tRCD of 3 for:

254FSB 1:1 = CPU 3048 MHZ/ DDR 508 MHZ @ 2.5-4-3-8

maybe something for PC4000 testing :)

[oldfart]

Thanks for the very through review. I've been fighting to make this point for some time. Yes, CPU speed is what really matters (I'll second the DUH!) here. Memory speed/timings do matter, but are a very minor tweak compared to CPU speed.

Your testing was with the same memory timings with both 1:1 and 5:4. If you were able to use memory that would allow tighter timings @ 5:4, you would see the small 1:1 ratio gain erased with better timings.

This also shows why I am so against SiSoft/Aida/Memtest benches. They do not represent real system performance. A PC is not a memory bandwidth tester. It is a system where mem BW is one part of the equation, and by itself, does not have a function. The amount mem BW adds to the equation is in reality, very little.

I still see people limiting a CPU overclock speed so that they dont have to <GASP!> use a 5:4 ratio and "cripple" their performance. They sometimes then spend a pile of $$ for new PC3700/4000 ram and get nothing in return for it. They have been misled by the PC3700./4000/MUST RUN 1:1/SiSoft score hype.

Again thanks! For awhile, I thought I was the only sane person out here in respect to this stuff. It's really not that hard. Just run a few benches and the results are very obvious.

[ThugsRook]

very nice :beer:

[SupermanCK]

Great Read...thx

[Thor]

Hmm, maybe I should do some

275fsb 1:1 (3-4-4-8)
275fsb 5:4 (3-4-4-8)
275fsb 5:4 (2.5-4-4-7)

comparisons.