FSB vs. memory speeds

[MistaTastyCakes]

When overclocking, which makes a larger impact on overall system performance - memory speed or the FSB? Would an FSB of 155 w/ 412mhz DDR be faster than an FSB of 150mhz with DDR running at 450, for instance?

 

[Duvie]

It depends on the apps...I think the 38mhz of ddr would not counteract the 5fsb say with your 2.4b which would be 90mhz.....



Look at my tested comparison...

IN 3dmark it was a toss up with 90mhz more cpu but 80mhz less ddr...

In besweet mp3 encoder which is obviously a cpu dependent program as mem speed has little effect...

IN divx encoding you see bandwidth does help and the lower speed cpu beat the higher speed cpu by 10 seconds in what would represent about 5% the length of a real movie....so 20x that and it would have beat it by 200sec or 3min 20sec in a full movie encoding..interpolating some numbers and it is clear it may have taken a 3.4ghz clock speed to equal the 3.24ghz with 450mhz ddr in that type of test...rember that 400+mhz ddr was also cas 2.5 with the most relaxed timings...

Now in most gaming apps (look at thugs past test) they tend to follow a similar line....

 

[MistaTastyCakes]

I did a bit of testing and from my results, 2.8/412mhz DDR scored 12554 on my system, while 2.7/450mhz DDR scored approx. 12476. Not a large difference. In Sandra tests, the two barely differed, but the 450mhz DDR edged out the 412mhz DDR, which seems right. I have no DivX encoding software or DVD's to rip, so I can't test my settings with that.

I think I'm going to stick at 2.7ghz/450mhz DDR. More bandwidth, a bit slower on my PCI bus since I have no AGP/PCI lock, and it's not but 1% slower in 3D apps according to my benchmarks. Thanks for the good ideas

Edit: I wonder how RAM timings would affect these numbers, if much at all? I don't know if it's the same for you Duvie, but I can't even POST at 400+mhz with aggressive timings - I'm using the same relaxed timings as you.

 

[Duvie]

I have memtested my stuff and I could run my memory at 410mhz ddr still with 2,3,3,5 and 2.6v...I needed to bump it to cas 2.5 shortly thereafter....

I could boot cas 2 well into the 420's but would get some errors and memtest had errors...My ram is only rated 2,3,3,6 at 370mhz ddr so I got to be happy it does what it does...

remember some boards are far more picky then others...This same stick of ram could only do 426mhz ddr on my epox 4bda mobo stable. It did 442mhz ddr at most conservative settings but I had restart issues....On my epox 4g4a I could only get 408-410mhz stable period...no boot any higher...

 

[lssanjose]

sudhian media has shown that running in asynch as opposed to synch offers no real benefits. I'd rather keep the clock signals in synch and so that the bus doesn't have to really "wait" on the memory.

 

[Duvie]

Originally posted by: lssanjose
sudhian media has shown that running in asynch as opposed to synch offers no real benefits. I'd rather keep the clock signals in synch and so that the bus doesn't have to really "wait" on the memory.

I guess my test shoed in divx multimedia app that is not the case...On a equal or limited bus like the athlons were often running memory higher then fsb is a waste since the athlon would have peaked its bandwidth, the p4's are different....With ddr the p4 is actually severly limited by memory bandwidth. therefore rasing ddr speeds even up to 450 in my case is still not enough to feed the bandwidth capabilities of the p4..At a 720mhz quad pumped fsb like I am running the bandwidth of the cpu would be liek 5.7gb/s and yet my 450mhz ddr only has a max theoretical bandwidth of ~3.6gb/s...

Amd 266fsb chip has a max bandwidth of 2.1gb/s so therefore anything over 133fsb or pc2100 is wasteful without raising the fsb proportionally...

Amd 333fsb chip has a max bandwidth of 2.7 gb/s so therefroe pc2700 or 166fsb is the limit there for supplying the amd chip....

 

[gooseman]

There is a very good article at overclockers.com on the front page about this very subject.

 

[Duvie]

here is the link....

Overclockers.com article!!!


he states a lot of what I said above in the first part of the article.....

But then he shows he has made a lot of ASSumptions that frankly tell me he has not really tested it out much....

And I qoute:

I'm seeing people who think that memory bandwidth is even more important than CPU MHz.
It isn't (outside of extreme situations, which we don't have here).
In real life, under normal circumstances, increasing the CPU MHzage by X% increases performance far, far more than increasing memory bandwidth by the same X%.
As a very rough guideline, increasing the CPU MHz X% will get you an overall performance benefit yield of about 60%. So a 10% increase in the CPU speed will get you an average 6% performance increase.
In comparison, increasing memory bandwidth X% will get you a overall performance benefit yield of about 10%. So a 10% increase in memory bandwidth will get you an average 1% performance increase.
Big difference, isn't it?
Increasing CPU speed is a major contributor to performance. Increasing memory bandwidth is a minor contributor to performance.


Hmm this moron hasn't seen my divx testing!!! Or thugsrooks old gaming tests....He most also throw out all of the advantages that rdram had back in its days over SCDDR mobos before DCDDR mobos....memory bandwidth makes much more difference then this moron wants to believe...What a bunch of FUD...


3DMARK2K1SE
2.4@3.20 (4:5) 445 = 8692
2.4@3.24 (4:5) 450 = 8719
2.4@3.29 (1:1) 366 = 8698
2.4@3.33 (1:1) 370 = 8725 ***best***

this doesn't fit his moronic theory as well....notice how a cpu clocked 2.7% faster only had a .06% faster score...virtual tie..according to that dipshite it should have been 60% of the 2.7% or 1.7%...should have had a score of 8867...also remember this is cas 2.5,3,3,7 versus the cas 2,3,3,6...

BESWEET Mp3 Encoder (min/sec)
2.4@3.20 (4:5) 445 = redoing
2.4@3.24 (4:5) 450 = 4min/7sec
2.4@3.29 (1:1) 366 = 4min/3sec
2.4@3.33 (1:1) 370 = 3min/59sec ***best***

This one is cpu dependent as 2.7% cpu increase resulted in 3.3% increase in time...still doesn't follow his 60% of cpu increase theory....I would think real world apps would qualify under normal real-life uses...

GKNOT 0.27 DIVX 5.02pro (min/sec)
2.4@3.20 (4:5) 445 = 4min/23sec
2.4@3.24 (4:5) 450 = 4min/19sec***best***
2.4@3.29 (1:1) 366 = 4min/33sec
2.4@3.33 (1:1) 370 = 4min/29sec

Hmm his lame thoery falls to pieces here even more...This is a real world app but it doesn't conform to his contrived lack of real data thesis.

a cpu of 2.7% increase actually was -3.9% in performance...How can that be!!!!!

again according to him that 3.33ghz cpu should have been 1.7% faster or 4min/14sec...even the cpu 133mhz slower beat the 3.33ghz by ~1.5%...



Guys like this Piss me off as they make bold statements without backing it up or name the apps where this is seen..I can go pull up old reviews with pc1066 i850e platforms with its superior bandwidth over SCDDR would beat same speed cpu with SCDDR by 3-8% in most apps....When I run at 450mhz ddr I actually score higher then rdram pc1066 system...

bandwidth for a p4 which with SCDDR is often severly limited leaves that room to make memory speed and susequent bandwidth have a lot of punch. With DCDDR it is obvios 1:1 is the only way to go as it matches p4's bandwidth perfectly....

IN this article he is talking about P4's so don't say he may have been talking about amd cpus....read it and you will see...

 

[lssanjose]

calling him a moron only disqualifies your statement . I mean you're only basing your results on ONE test. You're pretty caught up in the #s as far as data rates go. But it's more about clockspeed. Why do you think that moving to a 133MHz bus was a smart move by intel in regards to the p4? Sure RDRAM looks great but it only in reality moved @ 100Mhz the data multiplier is 4x but we shouldn't be caught up in that multiplier. Improving the bus speed is always better than the memory speed.

And remember a theory is a theory. It's not law. Laws are things that are established principles that can't change, theories on the other hand have to right to change.

 

[Davegod]

all those ^^ benchmarks look identical to me:

first set 28 points max differential, what 8700 average, i run 3dmark2001 four times straight after each other and get a much larger % differences than that, without touching a thing. ALL benchmarks have noise, statistical error both systematic and random, the benchmarks are there only to reduce it as far as possible, cannot emiminate it anything like to the extent that 0.3% means anything other than "well within statistical error". Repeat the benchmark 10 times and 0.3% still wouldnt pass the 95% confidence.

10 seconds out of a total of 259, is 4% and maybe worth looking at if the test has been repeated often enough, but i suspect the variance each time would be >4% anyway. A consistently higher score of at least a few % on every run (bar outliers) over a proper statistical sample and you can say ok. 4% on one run doesnt even imply its faster, never mind prove it.

 

[Duvie]

Originally posted by: Davegod
all those ^^ benchmarks look identical to me:

first set 28 points max differential, what 8700 average, i run 3dmark2001 four times straight after each other and get a much larger % differences than that, without touching a thing. ALL benchmarks have noise, statistical error both systematic and random, the benchmarks are there only to reduce it as far as possible, cannot emiminate it anything like to the extent that 0.3% means anything other than "well within statistical error". Repeat the benchmark 10 times and 0.3% still wouldnt pass the 95% confidence.

10 seconds out of a total of 259, is 4% and maybe worth looking at if the test has been repeated often enough, but i suspect the variance each time would be >4% anyway. A consistently higher score of at least a few % on every run (bar outliers) over a proper statistical sample and you can say ok. 4% on one run doesnt even imply its faster, never mind prove it.

All the divx test were ran multiple times(little to no flux) as well as 3dmark is ran 3 times...I get 10-20 opint fluctuations on my 3dmark scores...With divx I hav rean same movie encoding 3 times at same speed and the max difference in a 2hour encoding has been under 1 min...trust me as a person who has done a lot of them the divx programs do not fluctuate greatly...If I was running other programs like surfing the net or playing music I could slow it down but this is not the case. Each test was on a clean reboot and just that test ran....

If I had 4% fluctuations then in those 2 hour encodings I should have +/- 5min swings...I dont think so....At under 1 min I see less then 1% ad more like .4-.6%, which means 1-2sec max in that short test...

Far more scientific then an article that states no proof just statinng real world normal apps!!! Hmmm...Yes it is a theoruy but it lacks solid conclusion....

 

[Duvie]

Originally posted by: lssanjose
calling him a moron only disqualifies your statement . I mean you're only basing your results on ONE test. You're pretty caught up in the #s as far as data rates go. But it's more about clockspeed. Why do you think that moving to a 133MHz bus was a smart move by intel in regards to the p4? Sure RDRAM looks great but it only in reality moved @ 100Mhz the data multiplier is 4x but we shouldn't be caught up in that multiplier. Improving the bus speed is always better than the memory speed.

And remember a theory is a theory. It's not law. Laws are things that are established principles that can't change, theories on the other hand have to right to change.

Doens't disqualify anything as I ran real world test and he states nothing to prove his conclusions....

MOre then one test obviously if you read above...It is a cross-section of test....I also have an autocadd test I will throw in as well. I also did all of this testing back in the days of my 1.6a@2.736ghz w/ 342mhz ddr versus 1.6@2.64ghz w/ 440mhz ddr...They showed same performance increase with bandwidth in divx encodings....

I also cross-referenced this to Thugsrook's testing in which he looked at gaming apps....

Also...like a mentioned above if bandwidth is no great deal versus cpu speed then can we just erase everything rdram has shown us in relationship to SCDDR mobos????


The guy is one of many MORONS that print freely on the internet any conclusion they wish to make without any real proveable or duplicatable data...pure FUD...he is likely a guy who has a p4 that can oc well but his ram sucks a bit so he has convinced himself the raw mhz is better. remember the p4 at 533mhz fsb has a peak bandwidth of 4.2gb/s...that would take 530+mhz ddr to fill in SCDDR systems....When you oc the fsb you are raising that peak bandwidth even more...No SCDDR mobo can fill those requirements and thus more bandwidth to fill that will have real world benefits and more then the unproven % marks he claims...