Hercules 9700 Extreme Overclocking - BS?

[Mintmaster]

I saw a link to Digit-Life recently:

http://www.digit-life.com/articles2/radeon/r9700pro-oc.html

They get a 9700 PRO to 420 MHz core, 750 MHz mem. This is pretty incredible, but I guess it's believable.

However, if you look at their results, they get up to 60% increase in performance with this overclock. Since the GPU is overclocked by 29% and the memory by 21%, how is this possible? Wouldn't the maximum increase be 29% when GPU limited and 21% when memory limited, with all realistic non-CPU limited situations lying in between?

I don't see why this would happen. Do you think they made up the numbers? That's a shame, because I used to think Digit-Life made some decent articles. This puts a big dent in their credibility (alongside their insistance that the Radeon does rip-mapping).


BTW, sorry if this is in the wrong forum. Wasn't quite sure where it fits.

 

[Bigus Dickus]

That doesn't make a lot of sense. I would think scaling with frequency (core/mem going up proportionately) would be less than 1.0.

There are several cases where they do show a 30%+ increase in framerates, with less than a 30% increase in core/mem speeds.

It don't make sense. It's like chewbacca livingon Endor... why would he live there with a bunch of Ewoks? It don't make sense.

 

[Crusher]

A lot of their graphs show an exponential increase in performance with a linear increase of the core clock at max memory speeds. That's just silly. As for that 60% increase in UT2003 at 1600x1200x32 with 16xAF and 4xAA, I guess it's possible. It's only 60 fps, and they were running the demo on default quality settings, which IIRC is pretty fugly looking, especially the texture detail.

 

[Nagorak]

Maybe they raised them both to the point that they're not hitting some bottlenecks in each, thus allowing for a greater performance gain.

 

[jb]

Maybe see Anadtech Article where they get a R300 core to 400mhz for a comparison..

 

[WaltC]

I saw a link to Digit-Life recently:

http://www.digit-life.com/articles2/radeon/r9700pro-oc.html

They get a 9700 PRO to 420 MHz core, 750 MHz mem. This is pretty incredible, but I guess it's believable.

However, if you look at their results, they get up to 60% increase in performance with this overclock. Since the GPU is overclocked by 29% and the memory by 21%, how is this possible? Wouldn't the maximum increase be 29% when GPU limited and 21% when memory limited, with all realistic non-CPU limited situations lying in between?

I don't see why this would happen. Do you think they made up the numbers? That's a shame, because I used to think Digit-Life made some decent articles. This puts a big dent in their credibility (alongside their insistance that the Radeon does rip-mapping).


BTW, sorry if this is in the wrong forum. Wasn't quite sure where it fits.

I'm inclined to think it's BS [maybe, see below] because the best overclock I can do with my ATI-built 9700 Pro is ~350MHz core and ~335MHz ram before I start seeing artifacts. I have seen these figures repeated several times by other reviewers on the I'net and think my card is about average in that respect.

Granted, the Herc does have different cooling than the ATI stock cooling. And it could very well be that the Herc stock cooling is *much better*--I don't really know at this point. If the GPU cooling is that much better, 420MHz is not out of sight, I suppose. And the ram does have heatsinks whereas the ATI-built versions do not (and I think they work.)

As far as the performance differences--it depends as I would think that modes using FSAA and AF heavily might tend to show the largest percentage gains (as compared to straight no FSAA/AF modes) because in this instance the cpu is much less of a limiting factor. In fact, with current cpus the chief benefit to having so much fill rate and horsepower is the use of AF and FSAA as the cpu will be the bottleneck otherwise.

I haven't read the review but agreed that the claims as you present them do seem extreme.

 

[Randell]

WaltC - as you havent read the review - you will relaise its not the speed thats been obtained thast BS, its the gain being greater than the core/mem overclock.

The speeds claimed arent being queried due to the unique cooling methodology - putting the case outside in the russian winter of temps below freezing

 

[Entropy]

I'm inclined to think it's BS [maybe, see below] because the best overclock I can do with my ATI-built 9700 Pro is ~350MHz core and ~335MHz ram before I start seeing artifacts. I have seen these figures repeated several times by other reviewers on the I'net and think my card is about average in that respect.

Granted, the Herc does have different cooling than the ATI stock cooling. And it could very well be that the Herc stock cooling is *much better*--I don't really know at this point. If the GPU cooling is that much better, 420MHz is not out of sight, I suppose. And the ram does have heatsinks whereas the ATI-built versions do not (and I think they work.)

They place the entire system in a -3 degrees centigrade environment, (actually, they put outdoors in Russia. ), and use hefty fans to ventilate with the sub zero air.

Digit-life is generally caracterized by quite knowledgeable people writing articles about pretty interesting stuff using graphs that are confusing as hell. All the trends in that article make perfect sense though.

The extreme datapoint is with Ut2003 1600x1200 16xQuality 4xAA, and I agree that getting superlinear improvement with core clock (at 750 Hz memory clock) looks odd. I could explain small such effects, but can't really say if they apply here or not. I can say that the crew at digit-life are too knowledgeable to lie in an obvious manner. Read their articles. Compare. Judge.

Entropy

Edit: oops, Randell beat me to it

 

[Randell]

it happens

 

[WaltC]

WaltC - as you havent read the review - you will relaise its not the speed thats been obtained thast BS, its the gain being greater than the core/mem overclock.

The speeds claimed arent being queried due to the unique cooling methodology - putting the case outside in the russian winter of temps below freezing

Randell & Entropy,

Thanks guys! OK, now that you've refreshed my memory I did read that one!... I thought it was pretty funny at the end when he said that his monitor started acting up (frost bite?) and so they pulled the plug and went back in and got warmed up again (presumably)...

Yea, with these wacky guys anything's possible...

 

[Nupraptor]

Granted, the Herc does have different cooling than the ATI stock cooling. And it could very well be that the Herc stock cooling is *much better*--I don't really know at this point

It's supposed to be. One of the selling points of the Hercules card is that it's supposed to be a champ when it comes to overclocking. I know someone who got their's up to 372/702, so 420/750 doesn't seem too far fetched, especially given the temperature this one was in. There should almost certainly be visible artifacting at those speeds, though.

Cheers,
Nup

 

[Bigus Dickus]

Maybe they were using the 400Mhz version of the 9700Pro from Tyan.

I love rumors. I wonder how this Tyan rumor will play out.

 

[MuFu]

Maybe they were using the 400Mhz version of the 9700Pro from Tyan.

I love rumors. I wonder how this Tyan rumor will play out.

What rumour?! Ok, The Inquirer have FUDded it up a bit but Tyan have the board up on their site and [H] also have a quote (from Tyan, I guess) saying "try 400MHz". The default clock of the card is 325MHz.

I wonder what the "overclocking optimised" drivers do?

MuFu.

 

[MuFu]

Ah... having checked Rage3D I see what you mean. People jumping to conclusions again.

I noticed a lot of people figuring that if it had faster memory it would clock higher. Do you think that that would be the case? Personally, I believe the the bus itself is capping frequencies (due to coupling issues etc) so <2.8ns DDR may not clock any higher without bus tweaks. This bodes well for the 9700 (AIB) overclocking if it comes with 3.3ns BGA. It may well have almost the same memory clock ceiling as the 9700 Pro.

A switch to SSTL_18 would no doubt let them scale further.

MuFu.

 

[Bigus Dickus]

I agree, probably just confusion and rumor. However, Tyan does have the first board design that is a departure from the reference ATI board. Who is to say that they didn't tweak the bus for increased memory speeds?

 

[Mintmaster]

All the trends in that article make perfect sense though.

The extreme datapoint is with Ut2003 1600x1200 16xQuality 4xAA, and I agree that getting superlinear improvement with core clock (at 750 Hz memory clock) looks odd. I could explain small such effects, but can't really say if they apply here or not. I can say that the crew at digit-life are too knowledgeable to lie in an obvious manner. Read their articles. Compare. Judge.

Entropy

Edit: oops, Randell beat me to it

There are quite a few games where the improvement was greater than 30%, not just UT2003 @ 16x12 w/ 16xAF 4xAA. I don't see how these could be explained. Say you have a card that's 300/300. If you raise both clocks by 33% to 400/400 without changing memory latency settings, you will get 33% faster framerates. Each frame still takes the same number of GPU cycles to complete (if not CPU limited at all), so performance is simply proportional to the clock rate. When you start changing mem/core clock ratios, then things get a bit tricky, but even so your increase will never exceed the max of the two individual overclocks.

The only possible explaination that I can think of is that the high overclocking is producing such severe artifacts that the GPU is doing less work per frame, but this seems very unlikely.

As for the quality of Digit-Life, I agree. I wouldn't expect them to lie in such an obvious manner. They usually know what they're talking about to a much greater degree than other sites. I guess they thought they could pull a fast one here.

 

[Nagorak]

There are quite a few games where the improvement was greater than 30%, not just UT2003 @ 16x12 w/ 16xAF 4xAA. I don't see how these could be explained. Say you have a card that's 300/300. If you raise both clocks by 33% to 400/400 without changing memory latency settings, you will get 33% faster framerates. Each frame still takes the same number of GPU cycles to complete (if not CPU limited at all), so performance is simply proportional to the clock rate. When you start changing mem/core clock ratios, then things get a bit tricky, but even so your increase will never exceed the max of the two individual overclocks.

The only possible explaination that I can think of is that the high overclocking is producing such severe artifacts that the GPU is doing less work per frame, but this seems very unlikely.

As for the quality of Digit-Life, I agree. I wouldn't expect them to lie in such an obvious manner. They usually know what they're talking about to a much greater degree than other sites. I guess they thought they could pull a fast one here.

You're assuming everything is linked and it's not... What if Frame A is heavy on the core and chokes unless it's running at 400 MHz, while Frame B is heavy on memory and chokes unless the memory is running at 400 MHz. Then Frame C is equally heavy on both. Then D is just a light frame. So when the card is running at default it choke, choke, double choke, runs fine. Meanwhile once it passes over some threshold it just blasts through the first 3 frames with no problems. I mean this is all hypothetical, but it could be happening.

It doesn't seem impossible to me that they could be pushing the core/memory to a point where it overcomes some hurdle that's holding it back thus allowing better performance. It's also possible they tweaked other stuff on their system... I don't really see why they would just make up this data, but you can't really disprove it unless you run the tests yourself. Just because it's not a linear improvement doesn't mean it's fake. Although I have to admit it does seem kind of weird...just playing the devil's advocate here.

 

[Humus]

Scaling better than linearly is simply not possible.

 

[BoddoZerg]

You're assuming everything is linked and it's not... What if Frame A is heavy on the core and chokes unless it's running at 400 MHz, while Frame B is heavy on memory and chokes unless the memory is running at 400 MHz. Then Frame C is equally heavy on both. Then D is just a light frame. So when the card is running at default it choke, choke, double choke, runs fine. Meanwhile once it passes over some threshold it just blasts through the first 3 frames with no problems. I mean this is all hypothetical, but it could be happening.

Yeah, and hypothetically the NV30 could outperform R300 if it had guaranteed 2:1 lossless compression.

How does a frame "choke" at 325 Mhz and suddenly "blast through" at 400 Mhz?

 

[demalion]

Scaling better than linearly is simply not possible.

Well, if performance is drastically affected by synchronizing, isn't it? (I blame your complete lack of qualifiers for the sense of Deja Vu I have ).

It would depend on whether, for example, a cascade of delays could be resolved by increasing clock speeds for memory and/or VPU.

For instance:

A sequence occurrs where task A and B and C needed to be done "x" times to do task D.

And further, task A and B were relatively short but required lots of data, and C was relatively long but required little data.

If C tried to start at specific intervals ("t") longer than its time to run and failed until A and B were done reading data, and increasing memory speed allowed A and B to complete within that interval, C could be done in 2/3 the time (150% speed increase) even if the clock speed increase was less than that.

Further, if D could be done in another specific time interval ("u") in parallel to an A, B, and C for a different set of data being in progress, but a D for that new data could not be done until this D was completed, if u < t could be achieved by increasing VPU speed could have yet another cascade effect.

Now, it strikes me as reasonable to say this type of cascade is unlikely to occur over several benchmarks and clock values as they have in that article, but I don't think it is impossible for such a thing to occur with the VPU,CPU, AGP bus, and application/game encompassing such a wide array of possibilities in fitting into the above scenario when the operation of any of these things are a "black box".

and stop trying to victimize my wallet!