Parhelia at 3dCenter featuring FableMark!

[LittlePenny]

http://www.3dcenter.org/artikel/parhelia_theoretische_benchmarks/

I found the GF4 @220 comparison interesting as well.

 

[Geeforcer]

Why did they downclock the memory on GF4? It already has a substantial bandwidth disadvantage, and this would make it only worse. I thought that the point of the third column was to try and keep things as equal as possible.

 

[Dolemite]

Why did they downclock the memory on GF4?

Obviously just to make a clock-for-clock comparison. Hell, it already beats the Parhelia in almost every benchmark, underclocked or not, so why complain?

 

[Geeforcer]

Why did they downclock the memory on GF4?

Obviously just to make a clock-for-clock comparison. Hell, it already beats the Parhelia in almost every benchmark, underclocked or not, so why complain?

I am not complaining, I just wanted to point out that if the idea was to compare two chips under identical conditions they should have done the opposite, IMO if they wanted an effective "all things being equal" comparison, they should have underclocked the memory on the Parhelia to 162.5MHz. Right now, all their test reveals is how Parhelia perfroms relative to GF4 with the same clockspeed and half the bandwidth.

 

[Dolemite]

Well, I don't think they can give GF4 a 256-bit bus, or take the DDR up to 550 MHz , but its always going to be a sort of apples to oranges comparison anyway, so clock-for-clock still seems logical to me.

 

[Bjorn]

Well, I don't think they can give GF4 a 256-bit bus, or take the DDR up to 550 MHz , but its always going to be a sort of apples to oranges comparison anyway, so clock-for-clock still seems logical to me.

Hmm, doesn't seem logical to me at all.

 

[Edge]

Logical for any vertex, or pixel shader test, that involves many instructions.

 

[Bjorn]

Logical for any vertex, or pixel shader test, that involves many instructions.

Yep, the GPU clock might be ok (although not exactly that interesting imo) to downclock for those comparisons.

But, not the mem speed which is what we discussed here.

 

[Hyp-X]

The test proves that the GF4 is faster with a 128bit bus, than the Parhelia with its 256bit bus.

Kinda makes the 256bit bus pointless (not in general, but in this case).

Parhelia should score 880 in all single texturing fillrate test - and it doesn't.
I don't know if it's the drivers' fault - it's too early to tell.

I still think they should have done 8x2 instead of 4x4. That would (could) rule even at this low clock.

 

[Geeforcer]

Well, I don't think they can give GF4 a 256-bit bus, or take the DDR up to 550 MHz , but its always going to be a sort of apples to oranges comparison anyway, so clock-for-clock still seems logical to me.

But what it the point of this comparison then? Here is a hypothetical scenario: we have two cards, card A is clocked at 135/183MHz and has SDR RAM and card B is clocked at 120/150(300)MHz and has DDR RAM. If we were to use the same mythology as 3Dcenter did in their Parhelia article, we would have to downclock card A to 120/150MHz and compare the results.

Now here is the fun part: Card A is a special edition of Gefroce 1 SDR, while card B is a standard Gefroce 1 DDR. After running the tests we would find that is some of them card B would perform much better due to much greater bandwidth(not quite twice, considering DDR RAM inefficiency, but close). Using 3Dcenter's method, we would be lead to conclude that chip powering card B performs better clock-per-clock, a flawed conclusion considering the fact that GPUs on both cards are identical.

 

[CarstenS]

Being kind of a regular at 3D-Centers's Board, i've followed the trails of this "clock for clock" comparison almost from the beginning.

What Leonidas tried to find out here is, if Parhelia's performance is being limited by it's way to low pixel- (and if applied to modern games) even texelfillrate.

Furthermore, GF4 not only lowers it's fillrates with descending clockspeeds, but also the effectiveness of it's HSR implementation as fewer occluded pixels/polygons can be discarded.

Essentially, this is trying to act as a scenario where bandwidth is not of such utmost importance, as in 1600x1200x32 with maybe FSAA and AF turned on. In 1024x32 even a GF3 is limited mostly by it's core-clock.

 

[Bjorn]

Essentially, this is trying to act as a scenario where bandwidth is not of such utmost importance..

You say that you want a scenario that isn't bandwidh limited.
Ok, well, then lower the GPU clock speed and try to overclock the mem as mush as possible.

What do you gain by lowering the mem speed also in this case ?
(since you want a non bandwidh limited scenario)

 

[CarstenS]

8)
Yes, you're right. Sounds a bit odd to me too, but if you take a look at the Settings used for this test, i cannot see a single one, where even a regularly clocked GF4 Ti4400 (with both Speeds at 275MHz) would be Bandwidth limited.....

In fakt, i don't think, that it does make much of a difference in this aspect, if you leave gf4's mem-clock at default level, because in any case it does not come anywhere near the available bandwidth of Parhelia.

So, yes, you're right but it's practically irrelevant imho.

 

[Brent]

So to compare GF4 Ti 4600 and Parhelia equally on the memory bus what would the Parhelia's memory need to be if the Ti 4600 is at 325Mhz?

Alternatively if we downclocked the Ti 4600 to 275Mhz then the Parhelia would also need to be downclocked to 162.5Mhz?

I'm trying to learn what would be equal considering one has a 256bit bus and the other a 128bit bus, clock for clock, trying to equalize memory bandwidth.

 

[Geeforcer]

So to compare GF4 Ti 4600 and Parhelia equally on the memory bus what would the Parhelia's memory need to be if the Ti 4600 is at 325Mhz?

Alternatively if we downclocked the Ti 4600 to 275Mhz then the Parhelia would also need to be downclocked to 162.5Mhz?

I'm trying to learn what would be equal considering one has a 256bit bus and the other a 128bit bus, clock for clock, trying to equalize memory bandwidth.

Considering that Parhelia has twice the bus width, it should in theory have twice the bandwidth then a 128Bit card like GF4 if both have the same memory clock.

At default clockpeed GF4 bandwidth is

325 (frequency in MHz) * 128 (bus width in bits) * 2 (DDR memory is used) / 8 (to get result in bytes rather then bits) = 10400MB/s = 10.4GB/s

With Parhelia, we have 275 * 256 * 2 / 8 = 17.6GB/s

In order to get both cards at the same bandwidth, you would need to either overclock GF4 memory to 550MHz (1100 DDR), which will not happen, or to downclock Parhelia’s memory to 163MHz

 

[nggalai]

Hi there,

to clarify the question why the Parhelia's memory wasn't downclocked: Leo didn't find a tool that worked with the Parhelia board. The only way to go "proper" would then have been to raise the Ti4600's memory clock to 500MHz (which just doesn't do) or leave it as is.

So, Leo opted for downclocking the GF4's memory to check whether the 256bit bus has any clock-by-clock advantage over 128bits when comparing bandwidth-limiting benchmarks. I'm not too happy with that logic, either, but that's it.

ta,
-Sascha.rb