Radeon 9600 PRO Overclocking

[Nite_Hawk]

I disagree with both of you actually. Both the relative and absolute numbers are important.

An example where the relative speed is important would be a 2GHz chip that will only overclock to 2.1GHz. When you are talking about speeds in the GHz, a 100MHz overclock is fairly weak. It's very close to specification. On the other hand, a 50MHz chip that overclocks to 150MHz is very fast relative to it's specification. An example of why this matters, is say you have a 9700pro clocked at 325MHz, and a 5800ultra clocked at 500MHz. A benchmark shows that they perform the same at the baseline MHz. Both cards are capable of overclocking 50MHz. Based on the reference speeds, the 9700pro should gain MORE from a 50MHz overclock than the 5800ultra should.

Having said this, the abosolute numbers are important too. Imagine for instance, that you have a PIII 500MHz, and a PIII 1GHz chip. The 500MHz chip overclocks 100MHz and the PIII 1GHz overclocks 150MHz. THe 1GHz chip has a smaller relative overclock, but the actual performance increase is still greater than that of the PIII 500MHz even though it's relative performance is smaller. Personally I'd rather have a greater actual speed increase even if it was a smaller relative speed increase any day.

Nite_Hawk

 

[Saem]

Thats an interesting point. However i think that history shows that GPU's and CPU trends cannot be compared to each other.

This isn't anything to do with design, considerations of problem set --which is where CPUs and GPUs diverge significantly. Here we have processors which are experiencing increase in the rate at which they operate.

In an ideal world, if you increase the rate at which your processor runs, you see a linear increase in performance directly proportional to the percentage increase in rate. An absolute increase isn't worth much if that is a percentage increase of a few percent. The CPU comparision works perfectly. I'm not sure why you don't see that it does.

So when you OC two GPUs by the same proportion the performance increase will be roughly equal to how much they were OCed by. It just won't be 1:1 due to diminishing returns.

 

[Saem]

Having said this, the abosolute numbers are important too. Imagine for instance, that you have a PIII 500MHz, and a PIII 1GHz chip. The 500MHz chip overclocks 100MHz and the PIII 1GHz overclocks 150MHz. THe 1GHz chip has a smaller relative overclock, but the actual performance increase is still greater than that of the PIII 500MHz even though it's relative performance is smaller. Personally I'd rather have a greater actual speed increase even if it was a smaller relative speed increase any day.

Here you are comparing the same device and in the GPU example we're comparing two dramatically different designs.

Absolute is important within the same family, I agree with you there. Outside of the family, the only thing that matters is performance.

 

[Dave Baumann]

Evildeus: OK, I've had a short go with the 5600U and managed to get a stable overclock of 435/400. Here's a few scores with the 43.45 drivers:

                     350/350  435/400
RtCW (16x12) 49.9       60.4
3DM03 GT2     6.1         7.5
3DM03 PS2.0   9.2        11.2

 

[Evildeus]

Thanx dave 25%/15% increase, not bad. the performance increase is interesting also!

Nite_Hawk, sure i'll get some better performance, but which one is the more OCing? It's not because one OC by 500 MHz and the other by 100 that's it's not the latter which OC the most.

Sure the performance will depends on the initial performance, as you and i say, but that's not the point.

 

[mongoled]

Hello guys I have registered to say one thing.......

PLEASE, PLEASE, PLEASE, PLEASE, PLEASE, PLEASE.......

Rip of the punny heatsink and stick something bigger on that card!!

Pwitty pleaseeeeeeeeee....................

I really , really want to see the card perform with a better heatsink and am very defo not alone on this. Could you please put me out my misery and do that?

Please, desperation is the only word for my request.

I want to c what will happen to the clock speeds!!!!!!!!!!!

 

[WaltC]

There's quite a bit of illogic here when using relative numbers exclusively, without reference to the specific chip architectures. For instance, a Celeron is not in the class of the P4 architecturally, hence to say a 50% Celeron 300 overclock remotely compares relatively to a 7% 2GHz P4 overclock is to make an essentially meaningless statement. The two cpus are not comparable directly because of architectures. Likewise, it is equally unwise to do a relative overclock comparison of an RV350 and a GF4 Ti4200, since the architectures are much different and the RV350 has capabilities the GF4 gpu does not--exclusive of the overclock. Therefore to me it seems they cannot be accurately compared on a relative basis as to the percentage of overclock MHz.

What I think is the most interesting aspect "relative" to RV350 here with respect to its apparent ability to overclock so well is the idea that ATi is apparently getting excellent yields with RV350--yields which probably exceed the company's original expectations. This seems to underscore ATi's comments some months ago that the company was "having no problems" with it's .13 micron transition efforts.

MHz clocks and manufacturing process sizes are loose, general descriptors. Ultimately, the physical architecture of a given chip can have a dramatic impact on both.

 

[RussSchultz]

The ability to "overclock" really doesn't indicate yield.

All it indicates is there is plenty of margin in the critical path timing on the parts that pass test.

 

[Johnny Rotten]

There's quite a bit of illogic here when using relative numbers exclusively, without reference to the specific chip architectures. For instance, a Celeron is not in the class of the P4 architecturally, hence to say a 50% Celeron 300 overclock remotely compares relatively to a 7% 2GHz P4 overclock is to make an essentially meaningless statement. The two cpus are not comparable directly because of architectures.

Meaningless? Perhaps you're misunderstanding. The fact each architecture overclocks 150mhz is rather meaningless unto itself. Overclocking the celeron by 150mhz (50%) rather obviously provides a FAR more dramatic performance boost (compared to the its original performance) than overclocking a 2ghz P4 by 150mhz. The only thing being compared are the relative gains each cpu is making based off an overclock of an absolute number.

 

[Pete]

It can also mean a greater yield of high-performance VPU's, which they can brand as such and sell for mo' money. I suppose there's viable yield, and then there's speed-binning yield.

I now yield the floor to the relative vs. absolute sophists.

 

[WaltC]

The ability to "overclock" really doesn't indicate yield.

All it indicates is there is plenty of margin in the critical path timing on the parts that pass test.

I should have said "yield at the target MHz"--which I think it does indicate... If the target MHz for the chip was 500MHz, yields might not have been nearly as good. But I do see your point--just because a chip overclocks to xMHz does not mean it will validate at xMHz.

 

[demalion]

In response to the "sophistry"

The point being, relative numbers are what's interesting, not necessarily raw MHz.

That's right, absolute numbers never gives us some information, relative does.

I agree with Russ to an extent, but not with you at all.

Russ's point, as I understand it, is that relative numbers give us the indication of improvement (and that Russ is only concerned about that). While I think that is a silly thing to have as your only concern, it is understandable given what he does for a living. Poor Russ. (need that stick poking emoticon).

That doesn't make absolute numbers without meaning, since the context in which we use the result is absolute, not relative (we don't run a GF 4 Ti and a 9600 on the same machine at the same time for gaming, we pick one and are concerned for the performance we get out of it after all the other concerns are factored in).

The 9600 is a 4x1 chip with PS 2.0 shaders at 325 and 400 MHz.
The GF 4 Ti is a 4x2 chip family, PS 1.3 shaders, with 250, 275, and 300 MHz clock speeds.
The 9500 Pro is a 8x1 chip with PS 2.0 shaders at 275 MHz.

The features of the 9500 Pro and 9600 Pro are the same, so with the 9500 Pro starting out higher in absolute performance, the % overclock available to the 9500 Pro is important precisely because of the absolute numbers that result (that allow it to exceed the 9600 Pro overclocked performance mentioned)...to maintain consistency (in anything other than knocking the 9600 ) your argument would have it that a 20% overclock of the 9500 Pro would matter less than a 30% overclock of the 9600 Pro.

Why that doesn't make sense is exactly the same reason why the 9600 ~35-40% overclock is important and impressive: for its absolute effect on the final result.

Even ignoring the distinction between the 9600 Pro being shipped at the high end of the chip's offered frequency range, and the GF 4 Ti 4200 being delivered at the lowest clock frequency in the chip's range, the GF 4 Ti 4200 would need to overclock to 338-350 MHz to offer a similar relative gain.
Hmm...in the absolute result, we have PS 2.0 4x1 at 540-567 MHz with minimal cooling versus PS 1.3 4x2 at 338-350 MHz.
Also, a Ti 4600 would have to overclock to over 400 MHz for the same relative gain.
Now, if you maintain it is no reason to get excited because the % change is in the same range (when ignoring where each chip started in its shipping clock speed range), I think you're being silly, as you won't be gaming with the % change, but its result. Are you maintaining the % change is not noticeable?
If you choose not to get excited because the absolute numbers don't stack up against an overclocked 9500 Pro, I tend to agree, but I also think that doesn't matter to people who end up with a 9600 and then try to overclock. I also think that doesn't concide well with bringing up the Ti 4200.

Hellbinder, yes between a card that goes from 200 to 260 and another card that goes 500 to 600, the first has the best OC abilities, even if one gains only 60 and the other 100. Sorry if you can't understand that.

An illustration of what you are ignoring about architecture and the absolute results mattering: if the 500 card was 8x1 and the 200 card was 4x1, I'd say the 500 card starting off at ~5x the performance of the 200 MHz part is exciting and significant, and stretching that to ~6x the performance of the 200 MHz part is even more exciting. At the time when the 8x1 500 card didn't exist, the 260 overclock for the 200 card was exciting...after the 500 card came out, a new context was established. This is something, the final result for the user, your analysis seems to try and ignore (and I don't get why EDIT: I don't get why that makes sense).

Of course, the numbers for the 9600 aren't presenting that magnitude of difference, but even a 30% overclock for the 9600 Pro is a valid reason to be more excited than a 50% overclock for the Ti 4200, if you care about what you're getting out of the overclock instead of just the percentage of it.

Saem: am I poking a stick at you too? I don't think your "absolute performance doesn't matter outside of the same family" stipulation is valid to the people who will actually be using the cards.

 

[RussSchultz]

Sorry, I got drug into something I wasn't discussing.

I thought somebody was saying "hey, its such a good design because it overclocks by 150 mhz".

Which is well, ummm, just talk, especially if you're discussing pure overclocking/design headroom and "free bang for your buck" as a consumer.

A celeron 300 did the same, but that was a 50% increase.
A 486/25 that went to 33 was also a 50% increase, but only 12 mhz. The celeron 2.2 go up to about 3Ghz, but that's less than 50%

Which gives/gave you more bang for your buck?

p.s. what do I do that deserved "poor russ" and "since that's what he does for a living"? I write firmware for MP3 chips. I have no idea how overclocking relates to my job.

 

[martrox]

Sorry, I got drug into something I wasn't discussing..

Those hallucinogenic drugs will get you every time.......

 

[Evildeus]

Re: In response to the "sophistry"

demalion, i think you should re-read my quote:

Hellbinder, yes between a card that goes from 200 to 260 and another card that goes 500 to 600, the first has the best OC abilities, even if one gains only 60 and the other 100.

Where does i speak of performance, architecture, etc.? The first has the best OC ability, that doesn't mean that the gain of performance, in absolute numbers is greater, or the architecture is better :?
And then, you need to take into account the time of realease, the initial performance etc.

 

[demalion]

...

p.s. what do I do that deserved "poor russ" and "since that's what he does for a living"?

Well, the and the request for a poking stick emoticon was intended to convey that I was joking, and not actually disparaging you.

I write firmware for MP3 chips.

And are familiar and concerned with chip clock speeds from the designer's viewpoint, in contrast to the consumer's. Is that an insufficient link?

I have no idea how overclocking relates to my job.

Hmm...Ok. I mistook the amount of levity I could use in addressing you, my mistake.

 

[demalion]

Re: In response to the "sophistry"

demalion, i think you should re-read my quote:

Hellbinder, yes between a card that goes from 200 to 260 and another card that goes 500 to 600, the first has the best OC abilities, even if one gains only 60 and the other 100.

Where does i speak of performance, architecture, etc.?

Here is something I agreed with:

More 40% than 50% but anyway, impressive. Not enough against an OC 9500 pro unfortunately :?

Here is what I don't agree with:

I was looking @ the OC of 9500 pro cards, and finally i don't find the OC 9600 pro really impressive . Sure it's over 500 MHz, but 9500 pro can do as much as 50% more on the clock rate, so....

They are both better than anything nVidia makes......

Sorry but i don't think so. The card is OC by 30% on B3D or HFR, 40% by [H], and i can find plenty of GF4 Ti 4200 with 25-35% OC. So what's so impressive? Some figures:

http://www.hardware.fr/news/lire/17-04-2003/

Tonight with a bit of luck Dave will give us some 5600 U OCing.

You are clearly discounting the importance of the result to the consumer in the second post. Is there something unclear in my discussion? It clearly outlines what I have a problem with, and what I do not, in directly the terms you yourself discussed.

The first has the best OC ability, that doesn't mean that the gain of performance, in absolute numbers is greater, or the architecture is better :?

What are you asking? Look at the conflict in your application of the discussion of overclocking. You use absolute numbers for addressing the significance, then support it by disparaging the % overclock by an example that focuses on ignoring the absolute numbers.

And then, you need to take into account the time of realease, the initial performance etc.

Yes, you do. Didn't I mention that? Your argument (the one I'm disagreeing with) does not.

Which part of this did I fail to get across in my post?

 

[Hellbinder]

Where does i speak of performance, architecture, etc.? The first has the best OC ability, that doesn't mean that the gain of performance, in absolute numbers is greater, or the architecture is better
And then, you need to take into account the time of realease, the initial performance etc.

Here is the problem.. Unless i am misunderstanding something you are saying.

Card a 200mhz OC = 260mhz

Card B 500mhz oc = 600mhz

The idea that card A is a better overclocker than card B is simply not correct. The whole point is you simply cannot only talk raw percentages. CPU's are generally rated for a much higher core speed than they are released at. Because they are designed to scale over an entire year with little tweakage. GPU's are not designed to scale over an entire year. Regardless of the IHV in question the tollerances generally and demonstratably fall between 20-50mhz. That is going to be true regardless of % of the total Core speed. This has been the case since the invention of Overclocking GPUs. Occasionally a Chip comes along that is designed so well that it allows for even higher overclocks. The Ti 4200 is not an example of this. It is a scaled down version of a 300mhz core. The 9700 is not an example of this either, nor is the 9500 for the same reason. Again Regardless of percentages GPU's have a target clock and are designed thereof. Thus.. When a GPU comes along that overclocks 167mhz... It IS a big deal and it IS a bigger overclock than any to come before it. The percentages mean nothing. The argument of performance returns in this case is not as great becuase the 9600pro becomes almost immidiately bandwidth limmited. However.. That does not change the demonstratable fact that no GPU in the history of this industry.. has overclocked 167mhz over what it is rated at. Look at the GFFXU it actually may tie as a close secong for this OC title. It is rated at 400mhz, yet they are running it with not that extravegant a cooling solution.. ok its a little more that stock ..... at 500mhz also. It is able to achieve a total of about 130mhz OC. The little brother chips of the FX adhere to the traditional pattern of Overclocking.

The notion that the general % of an overclock should be the measure needs to stop. Chips like this are an aboration not the norm. As the clock rates go up even further, and average overclocks still remain in the 20-50mhz level.. The point that i make today will become ever more obvious.
Or will it be said.. this chip sux becuase its 50mhz OC is only 1%.

 

[RussSchultz]

I'm just going to have to disagree as to what defines a better "overclocker". But that's ok.

 

[demalion]

Russ, aside from being offended by my comments, does that comment refer to me or just Hellbinder?

Note that my post disagrees with both Evildeus and Hellbinder on specific points.