2.8C... PC3200 or PC4000 RAM?

I've seen this question pop up several times lately. With PC3200 you can run tighter timings, but the DDR speed is lower. The PC4000 is more expensive and timings are more relaxed, but DDR speed is greater. Ok, I ran a quick comparison of these two types of RAM in my 2.8C system overclocked to 3.5 gig. Everything was set identical except RAM timing (tighest I could do in both examples).

Corsair PC4000 Pro... 2X512... 2.5,4,3,6 timing

Sandra CPU ALU: 10766
Sandra memory bandwidth: 6060/6026
PCmark2002 CPU: 8695
PCmark2002 memory: 11276
PCmark2002 HDD: 1319

Kingston HyperX PC3200 (BH-5 chips)... 2X512... 2,2,2,5 timing

Sandra CPU ALU: 10455
Sandra memory bandwidth: 5546/5499
PCmark2002 CPU: 8686
PCmark2002 memory: 10623
PCmark2002 HDD: 1324

There is no doubt that DDR500 gives you far more memory bandwidth than DDR400. For the 5:4 ratio and tight timings to match the performance, you need significantly higher FSB. Hope this helps.

To prove my point about high FSB making a difference when using the 5:4 ratio, here is what my 2.4C looks like overclocked to 3.4 gig. It'll do 3.5 gig, but I was lazy and had these benchmarks handy. RAM is 2X512 Mushkin PC3500 black level 2 with 2,3,2,5 timings.

Sandra CPU ALU: 10722
Sandra memory bandwidth: 6166/6160
PCmark2002 CPU: 8436
PCmark2002 memory: 11450
PCmark2002 HDD: 1379

See? Now the high FSB starts taking over. You can probably imagine that at 3.5 gig, the 2.4C really starts dominating, especially in memory bandwidth. Trouble is, with my IC7 and a limit of 2.8V vdimm, I can't go much higher at the 5:4 memory ratio, even though it'll do 3.5 gig at the 3:2 ratio.

Hey batboy,

You need to run some benches like 2001se. They'll tell you more about system performance than Sandra and PCmark. Another thing you may want to try is putting that 2.8C in your AI7 and running the memory at 1/1 with tight timmings. To give you an example of the power of 1/1 with tight timmings. I turned on my A/C and benched my 2.8C at 5/4 and 277 fsb. My 2001se score was 24000. I scored 23900 at 260 1/1.

What are scores for you (for each) on Super Pi (1M), and 3DMark01 (per test score)?

pkrew i see your XT scored 24000 on p4 2.8 @ 3600 i suppose (fsb 260 you said)

my Barton 2.6 @ stock speeds did 17200 in 3dmark2001, is this s*** normal?

The 24000 score was at 3.88 using 5/4. I got 23900 at 260fsb 1/1 or 3.64. I take it that you mean your 9700pro was at stock. That seems like a real good score for a stock 9700pro

nope.. sorry i forgot to say, i borrowed a 9800xt, to see if it's worth the upgrade, and on the Barton i mentioned it did 17200

Thats not normal...true in 3dmark2003 it did 6500, but still

Well, we're hijacking batboys thread, but I'm not sure whats normal for a stock XT on your system. With everything set to performance I got about 21000 at 3.5. Do a search on the ORB for similar systems with an XT and see what scores they got.

Sandra is always happy to show extra bandwidth... try running some programs that are affected by latency, like 3dmark2k1.

I have an interesting memory quirk for you....

I had an IS7 running a 2.4C at 285fsb ram 3:2 (190mhz) 2-2-2-6
I would get about 40s in superpi 1M and I would hit 19000-19500 3dmark01.

Now I have an AI7 and with the memory voltages I can run the 2.4C at the same 285fsb with 5:4 ram (228mhz) 2-2-2-6 thats a 38mhz gain on the memory with all other settings the same...the result?

40s in superpi and I just ran 3dmark01...19300.

Now I feel really stupid.....every intel overclocker says that memory speed is king.....why the heck did I get absolutely nothing from adding 38mhz (76mhz DDR) to my mem???

I need to do A LOT more testing and Im not going to stop until something claims repsonsibility for this crap score...

pkrew said:

Hey batboy,

You need to run some benches like 2001se. They'll tell you more about system performance than Sandra and PCmark. Another thing you may want to try is putting that 2.8C in your AI7 and running the memory at 1/1 with tight timmings. To give you an example of the power of 1/1 with tight timmings. I turned on my A/C and benched my 2.8C at 5/4 and 277 fsb. My 2001se score was 24000. I scored 23900 at 260 1/1.

Click to expand...

Phil is spot on there, you've got to do some application testing before declaring a winner. You've only run one significant memory performance test, the PCMark2002 memory (and it's just a synthetic to be take with a grain of salt like all such tests). Phil's suggestion to use 3DM2001se is also spot on, and the other suggestion to use superPi is also a great idea. Q3A fps at low resolution also clearly shows the fastest system.

For reference, my machine turns 11,283 on PCM2k2 running 1:1 at a mere 380MHz. The clock speed we run the ram at just doesn't matter much as we have enough bandwidth to prevent degregation in the application performance at less than 400MHz.

Latency is the imporant thing to optimize, and only the fact that increased memory clock speed can improve memory latency (if the timings aren't largely compromised to achieve the extra clock) allows improvement from it. Whether you can use the full measure of GAT, your memory timings, and the mode of operation all dwarf the effect of the exact memory clock rate. 1:1 does posses a minor edge in latency, but does not always translate into a measureable application performance advantage. All else equal, it always will, but rarely is all else equal.

One thing that I did find out about 1/1 is the ability to use more aggressive GAT settings. I suspect that this may be mobo dependant. On the max3 I could set F1 at 1/1, but could use nothing but auto at 5/4. F1 was about a 2% advantage. On the AI7 however, I can set it to at least speed racer at 5/4 and many time keep the F1.

this was the most confusing thread I have ever read.

blah i still dont know what to do with my ram

Ok kids... I just got back home from a long weekend roadtrip.

I saw a few requests for Super Pi and 3Dmark2001 benchmarks. I don't have much time tonight to do the Super Pi benchies, but since I already had 3Dmark2001 scores from the 2.4C @ 3.4 gig with 5:4 ratio and 2.8C @ 3.4 gig with 1:1 ratio, it was no problem to just run the 2.8C @ 3.4 gig with the 5:4 ratio (I still had the 2.8C and Mushkin RAM in the test mule system).

I used a 9700 Pro at default speed and all default settings for these 3Dmark2001 tests.

2.4C @ 3.4 gig using 5:4 ratio with 2,3,2,5 timings (2X512 Mushkin PC3500 level 2) = 17056

2.8C @ 3.4 gig using 5:4 ratio with 2,2,2,5 timings (2X512 Mushkin PC3500 level 2) = 17166

2.8C @ 3.4 gig using 1:1 ratio with 2.5,4,3,6 timings (2X512 Corsair PC4000 Pro) = 17138

The two 2.8C scores are nearly a draw. Looks like if all you are worried about is 3Dmark2001 scores, either 1:1 or 5:4 ratio is about the same.

Since I now have an AI7 that allows higher vdimm, I'll try running these tests on that mobo within a day or two once I have it fully up and running. All of the above listed scores so far were using an IC7 with a max of 2.8v.

Ok... I made time and ran Super Pi 1M. Results are as follows:

2.8C @ 3.4 gig using 5:4 ratio with 2,2,2,5 timings (2X512 Mushkin PC3500 level 2) = 41 seconds

2.8C @ 3.4 gig using 1:1 ratio with 2.5,4,3,6 timings (2X512 Corsair PC4000 Pro) = 41 seconds

Nice job batboy, that's pretty much what I'd expect. The reviews that I've read show about a 100 point difference so you're right in the margin of error. I'm a little supprised the the 2.8 beat out the 2.4 considering that the 2.4 should have been running at a higher memory fsb. Maybe the timmings were the difference.

Maybe I should of used SPD or more relaxed timings at 1:1, not everyone can get CAS 2.5 out of their PC4000 (2.8C comparison). But, then again, not everyone can get 2,2,2,5 out of their 5:4 overclocks either.

I wonder if the 2.4C @ 3.4 gig score would of been higher if I could of ran it at 2,2,2,5 instead of 2,3,2,5?

You may be right, I'm trying to remember if the reviews I read used CAS 3 or not. As far as the timmings on the 2.4 you'll find out when you put it in your AI7.

Oops... wrong thread...