PC4000 versus PC3200 Results!!

There's been a lot of debating on whether PC4000 with loose memory timings is better than PC3200 with tight timings. I've decided to test this myself using my rig below. To make sure I only changed the memory speed and timings I tested the system as follows:

3.25 GHz CPU
250 MHz FSB
GAT settings all set to "Auto"

For 1:1 ratio, 250 MHz 2.5-4-4-7 settings (Geil PC4000 1 GB)

PC Mark : CPU - 7996, Memory - 11148
SiSandra Memory : Buffered - 5969/5998, Unbuffered - 4755/4760
AquaMark03 - 48,261
3DMark2001 - 19,797
3DMark2003 - 6,218
MPeg2 Encoding (Premiere 6.5) - 2:03
Prime95 (2000 interations) - 1:23

For 5:4 ratio, 200 MHz 2-2-2-5 settings (Corsair XMS3500 1 GB)

PC Mark : CPU - 8000, Memory - 10483
SiSandra Memory : Buffered - 5550/5550, Unbuffered - 4383/4287
AquaMark03 - 48,402
3DMark2001 - 19,943
3DMark2003 - 6,230
MPeg2 Encoding (Premiere 6.5) - 2:02
Prime95 (2000 interations) - 1:23

So what can we conclude?

1) Synthetic memory benchmarks show a clear advantage with the PC4000 memory.

2) Graphics benchmarks show a very slight to insignificant advantage with PC3200 memory.

3) CPU intensive benchmarks show no difference.

There it is folks, basically there is no difference in "real world" applications for either memory. What's more important is your CPU speed. If I run the same benchmarks at my normal speed of 3.45 GHz, I get real gains accross the board. So my advice to you is to get the cheapest memory that will run at one of these configurations and use the left over cash to buy the absolute fastest CPU you can afford.

However since I now own both of the tested memory, I have to say that the Geil PC4000 memory is pretty damn cool! I haven't tested above 260 yet since all Abit Max3 boards currently have a problem running higher voltages than 2.8 VDimm. As soon as I get this sorted out, I plan on testing again at 266 versus 214 memory. Probably no difference again but I like the idea of having 8.4 GB of memory bandwidth!

Peace....

Very interesting results. Even though running the memory at 1:1 results in higher memory bandwidth, you actually scored higher in Aquamark and both 3dmark's running the memory at 5:4 with the tighter timings. Thanks for taking the time to post these results, I'm sure a lot of people have been curious to see this type info.

I have an older game that responds dramatically to lower cas latencies, but only slightly to fsb and cpu speeds. I don't know how much this applies to modern games or programs, but you can bet it matters somewhere.

While I've never subscribed to the theory that cas latency is more important than total bandwidth, I will not hesitate to say that it makes a difference. And I also won't hesitate to say that in SOME applications, it does outweigh total bandwidth. It's kinda like seek time versus read speed for hard drives.

You will never come up with a definitive answer to this question on the forums. The only thing to do is to try different settings with certain games or applications and see which does better, then post the results. Then we can see which is better for a particular program, and we can make our choices according to what we do most.

And something like that would be difficult to do.

Thanks for sharing the results.

it would be interesting to see what the performance comparison would be at the higher speeds lets say 285-290 fsb speeds and would the pc3200 be able to keep up at 5/4 ratio and the better timings...

Basically, if you can keep the FSB at the same level as the FSB that you are comparing, then scores will be very similar to lokicat's using 5:4 and a tRCD-tRP of 2-2 compared to using 1:1 and a tRCD-tRP of 4-4.

Say you get your DDR500 up to DDR600, and use a FSB:Memory of 1:1 at 300MHz. Then you test your rig running your memory at DDR480 (say you are pumping it with 3+V or have miracle RAM ) with 2-2-2-6 timings and have it running 5:4 with an FSB of 300MHz. The performance comparison between the two systems (all other things being equal) would follow similar lines as lokicat's. That is the 1:1 (w/ 2.5-4-4) would have higher memory bandwidth, but performance in real-world applications/benchmarks would slightly lean towards the 5:4 (w/ 2-2-2) system.

Looking at pure performance numbers, there is no real winner. But add the price differance between PC4000 (high FSB) and PC3200 (low latency), then you tend to lean towards the PC3200 (low latency). Results like this have an impact on Intel-users with P4C (800 FSB) processors, as well as (future) Athlon 64 users.

woo neat

Set the GAT to Turbo, a, a, d, d and try again, the results may be different

Tried it but no go...

larva said:

Set the GAT to Turbo, a, a, d, d and try again, the results may be different

Click to expand...

Hangs after DMI verification.

More results at 260 FSB

Here's more data when I ran 260 FSB:

260 Memory 1:1 Ratio 3-4-4-8 Timings
Aquamark3 - 48,677
3DMark2001 - 20,267
3DMark2003 6258

208 Memory 5:4 Ratio 2-2-2-5 Timings
Aquamark3 - 48,699
3DMark2001 - 20,340
3DMark2003 6263

The differences between the two memory speeds are much closer than the tests I ran at 250 FSB. I suspect that the 1:1 ratio with loose timings meets or exceeds 5:4 ratio with tight timings at around 270 FSB. If I can only get my Max3 board working at more than 2.8 Vdimm I could test this theory.

...

omg...will this finally be then end of people saying "you are stupid...pc4x00 is faster...it has a higher # doesn't it?"...thank you for posting this. Nice layout...even the simplest of people will be able to read this info quite easily...many many links to be sent here...

Re: Tried it but no go...

lokicat said:

Hangs after DMI verification.

Click to expand...

The PC4000 won't allow GAT because it is um, less than impressive ram. The Corsair won't allow GAT because it is CH5. Put some BH5s in it and set GAT to F1 and your 5:4 configuration will extend its lead significantly.

The 1:1 configuration would win if it would tolerate GAT, even at the loose timings you have to run. But generally if ram is going to require loose timings it's going to balk at GAT also.

The real winner is to run 1:1, 2-2-2-5, and GAT on F1. You just need a higher cpu multiplier to make it come off. It's still the way to go even if your multiplier is too high, as mine is at 18. 17 would be perfect and allow you to exploit the cumulative advantages of 1:1 operation, tight timings, and a big fat load of GAT. As ram quality improves this will be possible at higher clock rates than the 400-440MHz currently achievable in DC with BH5s.

Most of the statements about bandwidth vs. latency reside at far too simplistic a level to say anything meaningful. Which is more important, latency or bandwidth? Well it all depends on the circumstances. For single channel chipsets the answer is almost always bandwidth, as they have excellent latency in any event but poor bandwidth. Obvously increasing bandwidth is more important in this application.

But 865/875 chipsets are an entirely different story. Firstly, they only achieve latency that approaches their single channel predecessor when run with tight timings. Secondly, they can match 845x's latency when run in 1:1 mode with tight timings. Thirdly, the can actually improve upon 845x's latency when they are run 1:1 at tight timings with PAT/GAT enabled. And fourthly, they don't lack for bandwidth no matter what clock speed you are running the memory, making the pursuit of better latency behavior more important.

Since our bandwidth in dual channel is sufficient to very nearly maximize system performance when running the ram at 400MHz, if it costs us latency performance to raise the bandwidth we lose, lose, lose. Since bandwidth is always in abundant supply, we need to work on optimizing latency in order to achieve overall improvements. Obvously running tight timings is the first step. If this requires going to the 5:4 ratio to achieve, do it. 1:1 is beneficial, in that it improves latency in itself, and PAT/GAT is more effective at 1:1 than any other ratio, further helping latency. But these effects are not as pronounced as getting the timings reduced, and if your ram will not tolerate GAT this further erodes the advantage of 1:1 operation.

aj aj aj .. dang this is heavy stuff .. makes me dizzy

Still great posts all ... keep wanting more

this stuff is like icecream .. could eat it all day if it wasn't for the headaches

IMO The problem comes when you put all that larva has said together with the varying multipliers of P4 and the varying performance of the CPU based on the FSB.
I can get my P4C to 330FSB in single channel mode, and although I am only running the P4 on an 8x multi for 2.64ghz, the Sandra cpu results are nearer to the 3ghz reference speed becuase the CPU becomes much more efficent at such a high FSB.
Really I want to be running in the region of 350FSB. You then have to think about running the RAM at a high speed too, but getting that high on 5:4 would not be easy, at least not at agressive timings to get DDR533.
Although understandably most people dont have chips with such a low multi, its hard to generalise based on so many variables...

Hmm, this is interesting. I have been comtemplating changing memory right now to get the 1/1 at 250 fsb/250 memory. Right now I'm running 5/4, 250 fsb-200 memory. It looks like now I should just stay where I'm at.

I'm running lower benchmarks though than posted above.

In 3DMark 01, I'm getting about 18,800.
3DMark 03, about the same as the starting above post.
SiSoft Sandra memory bandwidth, I'm getting about 4970
PCMark, I'm getting 9970 for the memory score.

And this is with my lower sig system compared to one above at 3.25 gig. I don't understand why I'm not getting the scores more in line with the above post. Any comments on this? Is it because of my 4 sticks?

On this topic I think to get a better 'real world performance' comparison though would be to give some comparison benchmarks running actual games.

Mike89 said:

Hmm, this is interesting. I have been comtemplating changing memory right now to get the 1/1 at 250 fsb/250 memory. Right now I'm running 5/4, 250 fsb-200 memory. It looks like now I should just stay where I'm at.

I'm running lower benchmarks though than posted above.

In 3DMark 01, I'm getting about 18,800.
3DMark 03, about the same as the starting above post.
SiSoft Sandra memory bandwidth, I'm getting about 9970.

And this is with my lower sig system compared to one above at 3.25 gig. I don't understand why I'm not getting the scores more in line with the above post. Any comments on this? Is it because of my 4 sticks?

On this topic I think to get a better 'real world performance' comparison though would be to give some comparison benchmarks running actual games.

Click to expand...

wow...9970 in s-sandra...that is the highest score ever recorded.

For your clock those scores are fine, and there is no reason for you to switch memory. Even if you had 1:1 crap that I don't like I wouldn't suggest switching memory. The difference might be kind of big in 3dmark, but you probably won't be able to notice the difference in your computer. If you run 2-2-2-5 5:4 you are getting better "real world performance" than if you bought crap that did 1:1 3-4-4-8.

btw...is that your highest oc?

wow...9970 in s-sandra...that is the highest score ever recorded

Click to expand...

Sorry about that, I screwed up. It was 9970 in PCmark02 (memory). Sisoft was 4970. (I went back up and edited that post)

The starting post got over 10,000 in both his comparisons (pcmark memory score), both with a 3.25 gig. Here I'm running at 3.5 gig and can't even break 10,000. And he's getting over 1000 more in 3dmark01 with either comparison, again running 3.25 gig. The only one I'm about the same in was the 3dmark03 score.

Dunno why my system is so far behind those.

Oh forgot to mention, my 4 sticks are running 2-3-3-7.

...

do you have cpc enabled, and what do you have all the other memory options set to?

What is cpc?

I have Hyperthreading enabled in BIOS. Dunno which other settings you are referring to for me to show. Lemme know which ones.

CPUZ showed PAT disabled, even though I have the Game Accelerator settings to Turbo-Auto-5-Disabled-Disabled. Windows won't boot with anything higher than Turbo.


I'm at home now (I was at work before) so I can now see my scores. I'm actually a bit lower in the PCmark scores even than I first said.

PCMark
CPU - 8278
Memory - 9954

...

can you put the disabled's to auto? 1 of those is the pat I think...someone needs to jump in here and help you out that has a max3. I am doing the best I can from the little bit I have seen from that mobo...sorry if I am not being much help.