INTEL - Synchronous is not important

[stevejst]

There is a common misconception in overclocking community that sychronuous bus and memory clock is extremely important for memory bandwidth. That is true in AMD case but not in Intel case.
Here is a little test you can easy replicate yourself if you are running a potent Intel board. This is on IS7-E using 2x256 GEIL Ultra Platinum PC 3500 at 2-3-3-7 timings:

Syncronuous 1:1 using 210/210
Sandra: 5052/5091

Asyncronuous 5:4 using 236/189
Sandra: 5089/5063

As you can see the bandwidth is about the same when the CPU bus increases 26 and memory bus decreases 21.
So there is a certain loss due to waiting times (2-3 MHz) but that is negligible. In fact you can easily ofset that loss by using more aggressive timings that lower memory clock makes available.

Similar tests you can perform with all the clocks and timings you can come up with and it will all reach the same conclusion. This is all on 2.4 GHz HT Pentium 4.

If you want performance you need to find your own soft spot. I like mine at 262/209.6 using 5:4 that gives me 3.144 GHz and memory bandwidth of 5733/5703 in Sandra at 2-3-2-6 timings. I can get close to 6000/6000 but it is not worth the stress on the CPU in my view.

 

[stevejst]

And don't let anybody tels you that timings are not important. They are very important.
As said 2-3-2-6 at 262/209.6 gives me 5733/5703. If I relax it to 2-4-4-7 (CAS the same!) the loss is already obvious 5608/5629. To ofset this you have to increase the clock to 265.
Relaxing CAS will make this go significantly lower.

 

[Jeff7181]

Have you tried setting tRas to 11? link to the discussion

 

[stevejst]

IS7-E does not offer that. In fact GEIL is very picky about timings and some relaxed timings are not possible. For example CAS 3 is not possible.
I am surprise about this on tRAS. My contention was that is merely a sum of CAS + tRCD + 1 or 2 (2 for safety) and plays no role for performance. But I'll try and see, I can get 8 max.

Edited:
Nope, I see no gain, in fact I lost a little but that is probably just a matter of a single test.

 

[Jeff7181]

The only benefit I noticed by changing tRas to 11 is that low FPS in games didn't drop as low as with it at 7. So I guess then "closing the book" explanation used in the article is accurate.

 

[Lonyo]

Have you run Sandra tests unbuffered? That might show more of a difference in the scores, Anand's article seemed to show that buffered scores weren't necessarily that indicative of bandwidth, IIRC.

 

[Jeff7181]

Have you tried some real benchmarks? Like, can you do a UT2k3 botmatch benchmark to see what the difference is in games?

Also, another thing I find interesting... increasing the bus from 210 to 236 increases the clock speed of a 2.4C from 2520 to 2832... and look at the tiny little increase in performance you gained on the CPU score. I am of course assuming you don't have an engineering sample processor where you can freely adjust the multiplier =)

How bout doing this... set your FSB to 200... bench it... then set it as high as you can while keeping your memory in sync with the FSB, then bench it again, and see how performance changes.

 

[stevejst]

Have you run Sandra tests unbuffered? That might show more of a difference in the scores, Anand's article seemed to show that buffered scores weren't necessarily that indicative of bandwidth, IIRC.

I run them buffered, as usual. Synthetic benches are never neccessarily the indication of the "real" performance but they are the only way to test the hardware irrespective of a particular usage. Whenever you use any particular application or game you will get results more skewed than on any synthetic bench.

 

[Jeff7181]

Steve... am I right in assuming you tested at about 2.5 and 2.8 Ghz to get those results? Or are you able to change the multiplier to keep about the same CPU speed throughout the tests?

 

[stevejst]

You cannot change Pentium multipliers.

Unbuffered:
210/210 - 4399/4348
236/189 - 3932/3847

You might have something there, but why would you want to test your system unoptimized?
Overclockers.com editorial says that Anand is making lemonade out of lemons.

 

[Jeff7181]

Well, the who idea about testing whether running in sync is better than running async is to test that ONLY. Of course a 2.8 Ghz CPU is gonna score better in every test than a 2.5 Ghz CPU. That fact that those two score the same in memory benchmarks when you use a RAM divider on the 2.8 Ghz setting tells me that the RAM divider hurts performance, all else being equal. Do you have any games you could benchmark with? Like the UT2k3 Demo and HardOCP's benchmark utility for it?

 

[stevejst]

Here you go, 3DMark2001SE (Radeon 8500, slightly overclocked, don't ask):

210/210 - 10,663
236/189 - 10,629

Hell, pretty much the same. It proves my point.

Of course, higher clock processor at 5:4, that is the whole point. That is why you would run asynchronuously. There is a certain loss on asynchronicity due to waiting times but that is negligible and easily ofset by higher CPU clock and better memory timings. Of course if you have memory that can follow CPU on 1:1 you would run it that way.

(By the way, my system runs GAT, game acceleration though I do not play games, it is the max optimization I get that way).

 

[Jeff7181]

So if you can't gain at least 300 Mhz by running async, it's not worth it. That's what your findings tell me. If you would, try it again 210/210, and then do it again at 210/168 and see how performance is effected with the same clock speed, but the RAM running on a divider. Maybe even set the resolution to 640x480 for this set of tests to eliminate the video card as a possible bottleneck.

 

[stevejst]

210/210, and then do it again at 210/168 and see how performance is effected with the same clock speed,

That would make no sense. I did not say it is better to run asynchronuously, it is certainly not!
But since if you have PC 3200 or PC 3500 that is the only way to increase the CPU clock, of course you would want to run it 5:4. The loss Pentium system would suffer is irrelevant compared to the gain due to aggressive memory timings and higher CPU clock.
5:4 will give you a chance to run 250 CPU and 200 memory. With PC 3200 you cannot do 250 on CPU with 1:1 because memory fails. You probably wouldn't be able to run 225/225 at 1:1. So 5:4 is the way to go.

 

[Jeff7181]

I'm not denying that you can recover the loss with increase CPU speed. But what I am saying is that Synchronous IS important. You're just introducing the same bottleneck that the 533 Mhz bus processors had, memory bandwidth. I'm not saying don't do that, I'm just saying you're creating a bottleneck, and that running async DOES effect the performance of a P4.

*EDIT* The purpose of running at 210/210 and 210/168 is to see how much of a bottleneck you're creating... if you loose 100 points in 3DMark, no big deal... if you loose 1000... that's a big deal.

 

[stevejst]

I am saying is that Synchronous IS important.

My numbers are showing (to me at least) that synchronocity is good to about 3 MHz on FSB. Talking about Pentium.
I find that irrelevant compared to timings and higher processor clock.

 

[Jeff7181]

Originally posted by: stevejst

I am saying is that Synchronous IS important.

My numbers are showing (to me at least) that synchronocity is good to about 3 MHz on FSB. Talking about Pentium.
I find that irrelevant compared to timings and higher processor clock.

So you're saying a 3.4 Ghz processor with RAM running at 200 Mhz will be faster than a 3 Ghz with a 1000 Mhz FSB and RAM in sync?

 

[stevejst]

I am saying that 238/190 is better than 210/210.

Subtract 3 MHz and you get (235+187)/2 = 211 ~ 210 = (210+210)/2.

Meaning - synchronicity is good for about 3 MHz. Say 3.5. And nothing more than that.

 

[Jeff7181]

I think your logic is flawed. Run the benchmarks, use percentages of the difference in performance.

BTW... why would you be running 238/190 if you have PC3500 RAM?

The way I see it is, at 210 in sync, you score 10,600... and when running a divider, you have to increase clock speed by about 12% just to regain what you lost by running async. So lets assume you can run your RAM at 217 like it's rated for. That would mean you can run your FSB at 271 with a 5:4 divider. Which would give you 3.25 Ghz. And if you ran at 217 sync, you'd be at 2.6 Ghz. So the question now is, how much better will it perform at 3.25, 5:4 than it performs at 2.6 1:1. And at what point will you "break even" when running a 5:4 divider vs. 2.6 @ 1:1. Based on my previous 12% calculation, at 2.9 Ghz with a 5:4 divider, you'd be getting the same performance as 2.6 @ 1:1. That may not be accurate though, because not every program needs a bunch of memory bandwidth, and may benefit more with higher CPU speeds, regardless of memory bandwidth.

 

[stevejst]

I think your logic is flawed.

It is pure math. I am mathematician so it is very easy for me.

BTW... why would you be running 238/190 if you have PC3500 RAM?

? For the sake of this test. I am actually running 262/209.6 at 2-3-2-6, I already said so! I cannot run 236/236 at 2-3-2-6. How can I compare then? That is the whole point! Are you stupid or what?

Whom am I talking to and wasting time? :disgust:

 

[Jeff7181]

Are you stupid or what?

Here you go with the insults again, are you incapable of having a civil conversation? Or are you just so damn arrogant that you get your shorts in a knot when someone questions you?

And BTW... you're not as smart as you think you are... the correct word is "who" the way you used it in the sentence, not "whom." "Whom" is used after a preposition, for example... "To whom am I talking to and wasting my time?"

 

[Tom]

Stevejst- I agree with your conclusions pretty much. For instance my system is using cheap Samsung pc2700 ram and a 3:2 ratio, but because it's running at 280fsb it gets benchmark scores better than a stock 3.2g system using a 1:1 ratio and pc3200.

And it's a helluva lot cheaper.

 

[Jeff7181]

Originally posted by: Dead Parrot Sketch
Stevejst- I agree with your conclusions pretty much. For instance my system is using cheap Samsung pc2700 ram and a 3:2 ratio, but because it's running at 280fsb it gets benchmark scores better than a stock 3.2g system using a 1:1 ratio and pc3200.

And it's a helluva lot cheaper.

A 3.36 Ghz processor is faster than a 3.2 Ghz processor... imagine that.

 

[WobbleWobble]

I'm going to have to agree with Jeff. The clock speed difference between your synchronous and asynchronous memory speeds makes up for the performance lost that an asynchronous bus/memory brings.

Yes, I understand that you're getting the same performance by bumping up your clock and running asynchronous. But what I think Jeff is saying is that your conclusion that "synchronous is not important" is wrong.

If you're getting similar performance by running asynchronously, why not just run it synchronously so you don't have to overclock your CPU as much?

And to draw a conclusion off of only two benchmarks is wrong. You need to run a wider variety of benchmarks to draw a conclusion. It's kind of like if you only look at two benchmarks where P4 beats an Athlon and conclude that a P4 is better, or vice versa.

 

[sman789]

try 3:2 please.....i've been advoiding massive OCing because of the "performance" loss