Does cpu OCing see diminishing returns?

[gentrinity]

I have this question. Does OCing a C2D see diminishing returns the more you push it. I mean, if you have it at 3.0GHz, and you bump it up to 3.4GHz, would you see the same incrimental performance boost you got from 2.4 to 2.8, or are there other factors like memory and stuff that will slowly start to give diminishing returns the higher you OClock? What about 3.6 to 4.0?

To throw specs name for this question, lets say youre going with a top of the line Asus, E6600, and 2GB Corsair XMS2 ddr2-800.

Thanks!!!!

 

[m25]

Generally, the more frequency increases, the more decreases performance, and that not only by overclocking but even stock operation. That means that for example, a 2.0GHz Athlon is something less than 10% faster than a 1.8GHz one. Of course, the more you get away from a base frequency, the more exponentially performance gain decreases.
There are special rules and math functions showing this (which of course I have not had the chance to memorize) but by benchmark results; a hell clocked P4 or CeleronD may finish SuperPI pretty fast but it's real world performance decreases miserably. This is the reason why intel didn't and still will not hit the 4.0 on stock (why it cancelled it's 4G P4), it just wasn't worth it.

 

[derek2006]

Y does it decrease?

 

[RichPLS]

Y does it decrease?

Just like speed in a car...
Going from 25mph (stock) to 50mph (overclock) doubles your speed via by 25mph or a 100% speed increase from stock...
but going from 50mph to 75mph, same 25mph bump, only returns a 50% increase from stock speed...
and going from 75mph to 100mph, same 25mph bump, now only returns a 33% increase in speed...

You get the idea now...

 

[dasickninja]

Jesus Christ Jack, everyone explained it in simple terms and in basic math, you had to go and break out the graph on us...
So does this all really matter? Thats my question.

 

[dasickninja]

Ahhh, agradeça para a resposta curta.
Finally the answer in a few words. I wonder when Joset will jump on this.

 

[ElMoIsEviL]

Jesus Christ Jack, everyone explained it in simple terms and in basic math, you had to go and break out the graph on us...
So does this all really matter? Thats my question.

What I know for a fact --- there are no FSB problems with the Core 2 Duo or the C2Q at 2.67 GHz, those are myths

It is one thing to say 'nope, won't be a problem', it is another to prove it

Let me put it another way.... all one hears from the Inq, HardOCP, Xbitlabs, is how the FSB is crippling the C2D.... this is a load of rubbish.

I can take the FSB all the way down to 400 MHz, clock the multiplier up to 20 (i.e. 5x faster than the FSB) and NO bottlenecks. I kick the FSB down from 1067 to 800 MHz, then multiply up the CPU to 3.4 GHz, guess what --- NO bottle necks. So what does this say --- the C2D does not consume much bandwidth, the FSB at 1067 is more than plenty, and if you back extrapolate, there is enough for a 2.67 Kentsfield ...

Jack

And there is a simple explanation for that which must be shared so that AMD fans who bang that drum understand that HTT/CSI is not useful for Core 2 thus far and won't be so long as we remain on a single socket. Whether Intel get's CSI or not is of no concern at this point. They WILL need it on the server side (multiple sockets) but on the desktop side it's not an issue assuming they continue using a shared caching system.

You see Core 2 Duo's (And Core 2 Quad's) use the cache for each core to talk to one another (Coherency). The Cache speed is faster then any HTT Link (as it runs at the speed of the processor) there is next to no latency and sure as hell no bandwidth problems. So for desktop level processors Intel has it down pat with their shared cache technology.

Now one would think that because Core 2 Quad (Kentsfield) has 2 pairs cores on a single piece of silicone and 2 Cache's (1 for each pair of Core 2's as essentially it's two Core 2 Duo's on a single package) that you would run into FSB bandwidth limitations as the two Core 2 Duo that comprise the Core 2 Quad need to talk to each other via the FSB. Well Tom's Hardware did a preview on the Core 2 Quadro and tests the processor at 1066MHz FSB as well as 1333MHz FSB and found there to be no performance increase. This indicated that the FSB is not starving the processors for bandwidth which again shows that for Desktop level processing CSI (or HTT if Intel adopted it) would not give the Core 2 Duo or Quadro any benefit at this point in time (4 cores). Perhaps scaling to 8 cores or two CPU sockets would require Intel's CSI technology (not perhaps, high probability that it would) but we're not there yet.

So to make a long story short, these beauties scale in a linear fashion. Even above 4GHz I have yet to see a decline in scaling performance. It's soo linear that it's damn scary, but this is to be expected as the higher I clock the processor.. the higher the 4MB L2 Cache get's clocked as it's on die running at full speed (1:1).

Hope this explains why Core 2 scales so well and suffer no FSB limitations.

 

[dasickninja]

And there is a simple explanation...

Simple you say?[/eye twitch]




This is going to attract the horde...... now.
....fire in the hole....

 

[ElMoIsEviL]

Yep, nice job.

Now if AMD can get their Cache Technology down pat as well with the K8L (specifically the L3 Shared Cache) we might get a rivalry going on that front.

AMD's Caching technology is like something out of 2001 (K7). Of course they made some pretty big changes from the K7, but the L1 Cache remained pretty much the same as did most of the features of the L2 Cache. It doesn't scale that well on bandwidth tests (specifically Sisoft Cache test), and even increasing the Cache has little to no impact on the performance of the processor (512KB vs 1MB). In fact 1MB K8's were pretty much useless.

I truly wonder what AMD have up their sleeves with K8L Cache wise.

 

[corvetteguy]

Yep, nice job. Once I see verified that the C2QX6700 is multiplier unlocked I will probably pick up one of those and repeat the experiments.

Tom's has already given us one data point (you pointed it out):
- a 2.67 GHz at 1067 MHz FSB performs the same as a 2.67 at 1333 MHz FSB -- so, at 1067 and 2.67 there is not FSB problems.

Jack

8O Pick up one eh....

LOL, you make it sound like going to the convienient store for milk

x6800 not good enough? :lol:

I'm just joking with ya cause i know your experimenting, but thats some expensive experiment.

 

[ElMoIsEviL]

And there is a simple explanation...

Simple you say?[/eye twitch]




This is going to attract the horde...... now.
....fire in the hole....

If the truth hurts then so be it. K8L will most likely be the most efficient architecture this time next year and I'll most likely be running a K8L 4x4+ rig as well.

If only "the horde", as you call them, were able to actually discuss these shortcomings without getting offended, perhaps then our discussions would remain civilized.

 

[ElMoIsEviL]

Yep, nice job. Once I see verified that the C2QX6700 is multiplier unlocked I will probably pick up one of those and repeat the experiments.

Tom's has already given us one data point (you pointed it out):
- a 2.67 GHz at 1067 MHz FSB performs the same as a 2.67 at 1333 MHz FSB -- so, at 1067 and 2.67 there is not FSB problems.

Jack

I feel that the 1066MHz FSB will probably start getting strained around 3.2GHz. This is probably why Intel have already moved there Extreme Edition versions over to the 1333MHz FSB as the same platform (perhaps a new motherboard due to voltage changes from 65nm to 45nm) will power their 45nm 3.83GHz models this time next year.

 

[RichPLS]

So you are adverse to sinning bitchez? :?:

 

[ElMoIsEviL]

A little advice in future, get your head out of Cinebench.
It's as bad&good as Bungholiomark and SuperPi(non-linear, see sticky).

It depends greatly on scenario. Not every benchmark increases linearly or let alone any increase at all. e.g. Oblivion at playable resolution--a GPU limited game

Games are not good indicators of CPU performance (as you pointed out GPU becoming the bottleneck). Cinebench seems to be a good indicator of scaling though (as are most Video Encoding applications).

Super PI relies too much on memory timings, but if you can maintain the same timings all the way through your experiment Jumpingjack (would be hard to do seeing as you're more then likely overclocking the ram) you'd get a pretty good idea of CPU scaling.

With Super PI it must all be done using the exact same timings, BIOS configuration, O/S, drivers, apps and system components.

It's quite the sensitive benchmark.

 

[endyen]

I've heard that the max speed of the FSB is 1666. Is this true?

 

[ElMoIsEviL]

I've heard that the max speed of the FSB is 1666. Is this true?

Well, we've been told for years now that they had reached a maximum speed... but we have yet to see that occur.

I believe Intel's last FSB speed will probably be the 1333MHz FSB, after which they will transition over to CSI. Only if CSI is late do I expect to see a 1600MHz FSB (400MHz Quad Pumped).

 

[sailer]

While no doubt, as Jumpingjack has shown in his graphs, overclocking can remain linear in its gains, at least up to some point , maybe the cpu melting, there are other considerations. So far, at least, the advantages of 16x lanes for the PCI-E graphics card have not been utilized, so that 16x boards aren't all that better than 8x boards. Then there's the other chokepoints, like how fast hard drives load to the ram, etc. The gains by overclocking the cpu may show in 3Dmark06 or something, but get down to playing a game and framerates often are ruled more by the graphics card and how well it comunicates with everything than how fast the cpu is.

Point being, that our computers are a package, and if the package isn't in synche with itself, you can have a blazing fast cpu and not see any significant change in performance. That said, as Jack showed in his graph, the C2D is one heck of a good cpu. Now if only everything else can get so good. Of course, this is just my humble opinion.

 

[randomizer]

I wonder if we will see significantly large caches in the near future (not 70 years) like a few hundred Mb of more. The cpu would have little reason to access the memory for much of the time (unless we have programs that take up ridiculously large amounts of memory, which we will coz manufacturers want us all to buy more RAM :lol: )

 

[randomizer]

Yea, but hey... if you own a multi-billion dollar corporation bent on global (computer market) domination, why not brag that you have 256Mb on-chip cache (I so wanted to type 1Gb, but I know that you arent going to fit anywhere near even 256Mb on a single die - unless you start making some fat processors...)

 

[randomizer]

What is world?

 

[mitch074]

From small tests I did with superPI, I may say that it's pretty much a memory bandwidth benchmark - I get the same results on a Sempron at 1.6 GHz and the same CPU at 2.0 GHz with locked RAM speed of 400 MHz (all timings being equals and HT link running at 800 MHz)

To test a CPU's throughput I usually launch two DVD to DivX (using Avisynth and Virtualdub) plus 2-4 AC3 to Ogg conversions (this maxes a dual core at 2x100%). In such a case, performances don't increase linearly it I don't increase memory and bus speed along with CPU clock - but then I don't have a C2D to test, so I'd be interested to get some returns from someone who has such a setup.

 

[mitch074]

Well then, I guess I would try to start a few mencoder and oggenc processes on my Linux OS - in command line, init 2 mode (multitasking, no extra services), with a SMP, non preemptible, 64-bit, no debug code, kernel - it would be the best test environment I could get.

How's that?

 

[gentrinity]

Wow, I started a nice little gem of a post here. Lots of great info.

To what you said Wusy, im a noob to all this, I just recently started doing research about a month or so back. Sorry if Im not too familiar with the lingo yet. But at least people understood what I meant. I wanted to know if processors did linear scaling in performance, and the main area of interest for me is video so I guess from that its a good idea to overclock for video and stuff.

I do ask the following question. How much would it cost me to get a E6600 up to 4.0GHz. I plan on starting out soft with a Scythe cooler or something of the sort to try and get 3.0GHz stable, and if all goes well, Ill look to spend some more money on better cooling to try and reach 4.0GHz, but what type of cooling would I be looking at and how much?

Cause Im not sure whether water cooling can keep temps stable at those speeds, especially since im doing rendering which is in itself a pretty arduous task for a CPU, and the fact that I cant have any artifacts appearing or anything of the sort. Stability is a must.

Thanks to everyone for contributing so well to my thread!!!!

EDIT: Not an E6400, an E6600

 

[randomizer]

How much would it cost me to get a E6600 up to 4.0GHz.

Probably enough to make up for the difference between an e6600 and an e6700, or maybe even an x6800

 

[buzz_X]

Hi gentrinity,

Yep, good post. My experience with OCing is quite limited but widely practical. A few years ago when the Athlon XP Mobile was the rage, I ordered two 2600's and they came with a very good stepping.

I wanted to build 2 computers:

1) General home computer (the one I'm using right now)

2) A video encoding screamer, since I wanted to encode my Divx videos at or near real time.

The general home computer was given a "sweet spot" of 12.5 x 200 FSB @ 1.75 V during winter months, and 12 X 200 FSB @ 1.67 V during summer months.

The video screamer is used only to encode video, and it runs with these settings: 11.5 X 237 FSB @ 1.92 V

Both the above computers run on air cooling only.

The only extra money they cost was in getting a good quality large copper heat sink and fan. I spent $25 on the general computer, and about $55 on the video screamer.

The most costly thing was the time spent on the video screamer, all the time it took to figure out the best settings for the motherboard and memory. However, 3 years ago I encoded my first 2 pass Divx movie in a little less than real time (up from a 450 Mhz PIII that took a couple of days !!!)

Yeah, I know both my little computers are starting to become dated. Yet and still, they were built for very little and they are still running well. Most important: The are still very pleasing to me!

So, yeah, my experience with OCing has been rewarding and productive because I new what I wanted.