5850 just as fast as a 5870?

[AzN]

I've read this whole thread. You're the only one here rambling on about things after obviously being proven wrong by evidence provided by others (you've barely provided any). If these are "hard facts" to you, that is a big problem.

I can't help you see things differently.

1. There is no proof that drivers are the culprit other than some guys within the hardware community saying so and mass pretty much follow the leader i suppose
2. 4870x2 still beats 5870 with lower core frequency but more bandwidth
3. 5850 overclocked to match 5870 core and more bandwidth beats 5870

 

[Daedalus685]

With no proof that drivers are the culprit I might add. Non bandwidth hardware limitation like? When only thing 4870x2 has over 5870 is bandwidth it's the most likely limitation of the 5870.

I don't understand how you can claim the 5870 is close enough to the 4870 that drivers should be identical, given the same pattern repeated from 2000 to 3000, and 3000 to 4000, all far more similar architectures than 4000 to 5000.

Could you please explain why dx11 addtions can not possibly change shader for shader scaling? And why drivers could not be broken given how we have seen a lot of performance jumps over the 2000, 3000, and 4000 series, all very closely realted cards?

Again, you insist the minimums should be included. These are instantaneous values, what relevance do they have? How do you recomend the data be processed? You say yourself that the minimums are pointless becasue of harddrive seeks, should reviewers manually remove data points they think are outliers? Does not a plot of the raw data, without any filtering, tell you more than a minimum value without pointing out whether it was filtered or not?

 

[Meaker10]

When memory bandwidth is limited, it not only causes drops in the minimum frame rate, it lowers the AVERAGE aswell, you would see some games have slight drops, and others just tank as you turn the details up.

In this case, overclocking the memory would increase performance significantly, we are not seeing that here.

 

[thilanliyan]

2. 4870x2 still beats 5870 with lower core frequency but more bandwidth

Isn't it only in PR slides and retail boxes that the 4870x2 has more ram and bandwidth than the 5870? IIRC bandwidth to each GPU is still the same as a single 4870.

 

[Daedalus685]

1. There is no proof that drivers are the culprit other than some guys within the hardware community saying so and mass pretty much follow the leader i suppose

No, there is no proof that it is drivers. However, something is effecting shader for shader scaling. Given that increasing the Bus barely increases the performance it is likely that this is not due to bandwidth. Sure it could be, but it seems far less likely than an inherant gpu difference, or bad drivers absed on the evidence we have.

2. 4870x2 still beats 5870 with lower core frequency but more bandwidth

They are not the same card, the specs may look similar but tehy are not the same. You have to understand that the comparison is not valid. Whether much was changed doesn't matter. The fact is that some of the core was changed to accomodate teh new features of the 5000 series. How are you so sure none of these can ahve any impact on performance. In fact, based on teh 5000 series it looks like the architecture of addign dx11 hurts shader scaling across the board. Every 5000 series part performs slower than an equivalent 4000 part.

3. 5850 overclocked to match 5870 core and more bandwidth beats 5870

This is the topic at hand. While I won't argue that the 5870 coudl be held back by bandwidth, clearly it could, what makes this any less likely to be a driver issue? Why when increasing the memory bandwidth on the 5870 does the performance of the 5850 still fall far too close?

 

[Daedalus685]

Isn't it only in PR slides and retail boxes that the 4870x2 has more ram and bandwidth than the 5870? IIRC bandwidth to each GPU is still the same as a single 4870.

AZN does not understand that useful bandwidth in a crossfire configuration is not equal to double what you have on each card. While teh 4870x2 still should have more useful bandwidth than a 5870, it is not all taht much more.. but would depend on how optimized the game, and crossfire itself, is.

 

[AzN]

I don't understand how you can claim the 5870 is close enough to the 4870 that drivers should be identical, given the same pattern repeated from 2000 to 3000, and 3000 to 4000, all far more similar architectures than 4000 to 5000.

Could you please explain why dx11 addtions can not possibly change shader for shader scaling? And why drivers could not be broken given how we have seen a lot of performance jumps over the 2000, 3000, and 4000 series, all very closely realted cards?

Again, you insist the minimums should be included. These are instantaneous values, what relevance do they have? How do you recomend the data be processed? You say yourself that the minimums are pointless becasue of harddrive seeks, should reviewers manually remove data points they think are outliers? Does not a plot of the raw data, without any filtering, tell you more than a minimum value without pointing out whether it was filtered or not?

We already went through this yesterday. There is slight upgrade on Cypress mostly prefetching (mostly artificial) and is unchanged from RV770 architecture.

Since most of these games that are tested are dx10 and lower why would dx11 additions apply here? It's not like tessellation or dx11 path is being used on a dx10 game.

I did not say minimum fps are pointless. I said it's most important aspect of the game. What I said was when testing minimum fps they should get rid of margin of errors like hard drive seeks and testing the same scenario multiple times is a must.

 

[AzN]

When memory bandwidth is limited, it not only causes drops in the minimum frame rate, it lowers the AVERAGE aswell, you would see some games have slight drops, and others just tank as you turn the details up.

In this case, overclocking the memory would increase performance significantly, we are not seeing that here.

lowers average frame rates slightly yes but minimum frame rates is where the bandwidth has the most effect on. It's the same when you raise resolution and use AA with more pixels applied. you need more bandwidth to peak those pixels.

example:
Let say a game has multiple scenes.

70% in confined space = Advantage to the core SP
20% semi open space = Advantage to the core SP and little effects of the bandwidth
10% very wide open space with massive graphics detail = advantage to bandwidth

You take an average from that which makes core the most difference in average frames. But in those limited situations that's where the bandwidth comes in.

 

[Daedalus685]

We already went through this yesterday. There is slight upgrade on Cypress mostly prefetching (mostly artificial) and is unchanged from RV770 architecture.

Since most of these games that are tested are dx10 and lower why would dx11 additions apply here? It's not like tessellation or dx11 path is being used on a dx10 game.

I did not say minimum fps are pointless. I said it's most important aspect of the game. What I said was when testing minimum fps they should get rid of margin of errors like hard drive seeks and testing the same scenario multiple times is a must.

But the hardware is there for dx11 whether you use it or not. However you want to think about it, a 4870 is not identical to a 5770. There are changes, period. I don't know how much these changes come into play.. but you can't just dismiss them as important because you feel like it. They are jsut as likely a cause for the slowdowns as any other change, and directly effect how the drivers are made.

No, for whatever reason you thinnk minimum is important. Yet you also believe that it is not important in situations where the data is not modified. What poitns shoudl they remove, what points should they leave? Shoudl they just use a running average over, say, 10 points? Or shoudl they go through and simply delete all the points that yuou don't seem to like.

I've already explained how minimum fps is an instanteous value. For it to mean anything it has to be presented in some way other than that. How should it be shown? An erratic plot may have a lower minimum FPS than another, yet the integral of it may very well be higher.

 

[AzN]

Isn't it only in PR slides and retail boxes that the 4870x2 has more ram and bandwidth than the 5870? IIRC bandwidth to each GPU is still the same as a single 4870.

PR slides or not 4870x2 with PR bandwidth is beating 5870 with lower bandwidth.

 

[Daedalus685]

lowers average frame rates slightly yes but minimum frame rates is where the bandwidth has the most effect on. It's the same when you raise resolution and use AA with more pixels applied. you need more bandwidth to peak those pixels.

example:
Let say a game has multiple scenes.

70% in confined space = Advantage to the core SP
20% semi open space = Advantage to the core SP and little effects of the bandwidth
10% very wide open space with massive graphics detail = advantage to bandwidth

You take an average from that which makes core the most difference in average frames. But in those limited situations that's where the bandwidth comes in.

Those situations are not "minimums" those are sustained minimums. The value reported in any review site is not the sustained minimum, it is the instantaneous minimum. They are completely different things. What you want reviews to include seem to be the minimum of the averages over various time blocks, or the integral of the time blocks. Not the minimum.

 

[AzN]

But the hardware is there for dx11 whether you use it or not. However you want to think about it, a 4870 is not identical to a 5770. There are changes, period. I don't know how much these changes come into play.. but you can't just dismiss them as important because you feel like it. They are jsut as likely a cause for the slowdowns as any other change, and directly effect how the drivers are made.

No, for whatever reason you thinnk minimum is important. Yet you also believe that it is not important in situations where the data is not modified. What poitns shoudl they remove, what points should they leave? Shoudl they just use a running average over, say, 10 points? Or shoudl they go through and simply delete all the points that yuou don't seem to like.

I've already explained how minimum fps is an instanteous value. For it to mean anything it has to be presented in some way other than that. How should it be shown? An erratic plot may have a lower minimum FPS than another, yet the integral of it may very well be higher.

You do know RV770 has tessellation as well. Dx11 are nothing more than instructions added to the core much like SSE4. Do you see E8400 with SSE4 support getting beat by E6600 with SSE3 enabled application at the same clocks?

 

[Daedalus685]

You do know RV770 has tessellation as well. Dx11 are nothing more than instructions added to the core much like SSE4. Do you see E8400 with SSE4 support getting beat by E6600 with SSE3 enabled application at the same clocks?

This is not an e8400.. Do you see a pineapple being eaten by many dinosaurs?

The 5870 is different than the 4870... I know that dx11 is relatively minor, but somewhere along the way something changed. How can you be so sure that this change did not make any difference in games? You are being totally illogical. Bias does seem to play a big role in one's ability to think clearly..

I am well aware of teh tesselation on the 4870.. Which is very different from the 5870's... Unlike others in this thread I am not pretending to know how every transistor in the 5870 works.. since you seem to, could you explain to me exactly why dx11 modifications, can not possibly have effected performance at all?

 

[AzN]

No, there is no proof that it is drivers. However, something is effecting shader for shader scaling. Given that increasing the Bus barely increases the performance it is likely that this is not due to bandwidth. Sure it could be, but it seems far less likely than an inherant gpu difference, or bad drivers absed on the evidence we have.

http://techreport.com/articles.x/17986/2

These are simply theoretical peak values, and they are in some cases quite academic. You're rarely going to hit the peak pixel fill rate on one of these cards, for instance, since memory bandwidth will likely limit you first. The GeForces probably won't ever reach their peak dual-issue shader FLOPS numbers, and I doubt the Radeons will achieve more than 80% of their peak compute capacity when executing pixel shaders.

Perfect example of the bandwidth limiting 5870. I've shown you this multiple times only for for you to graze right by without you answering any of this.

They are not the same card, the specs may look similar but tehy are not the same. You have to understand that the comparison is not valid. Whether much was changed doesn't matter. The fact is that some of the core was changed to accomodate teh new features of the 5000 series. How are you so sure none of these can ahve any impact on performance. In fact, based on teh 5000 series it looks like the architecture of addign dx11 hurts shader scaling across the board. Every 5000 series part performs slower than an equivalent 4000 part.

However you want to perceive is fine but by reading Cypress articles there's nothing more than prefetching difference, bandwidth saving techniques, etc. which are nothing more than PR marketing sham. SP works exactly the same with superscalar with 5 ALU's. Updated tessellation unit and added dx11 path.

This is the topic at hand. While I won't argue that the 5870 coudl be held back by bandwidth, clearly it could, what makes this any less likely to be a driver issue? Why when increasing the memory bandwidth on the 5870 does the performance of the 5850 still fall far too close?

5870 is already a fast card. More bandwidth would have been nice. not saying 5870 is an unbalanced card of any sort but with more bandwidth it would outperform 4870x2 and more inclined with GTX295. You guys leaving out bandwidth and blaming everything on drivers like some magic out of hat that is somehow going to get rid of the bandwidth limitation.

 

[AzN]

This is not an e8400.. Do you see a pineapple being eaten by many dinosaurs?

The 5870 is different than the 4870... I know that dx11 is relatively minor, but somewhere along the way something changed. How can you be so sure that this change did not make any difference in games? You are being totally illogical. Bias does seem to play a big role in one's ability to think clearly..

I am well aware of teh tesselation on the 4870.. Which is very different from the 5870's... Unlike others in this thread I am not pretending to know how every transistor in the 5870 works.. since you seem to, could you explain to me exactly why dx11 modifications, can not possibly have effected performance at all?

Comparing E6600 to E8400 and RV770 to cypress is within reason not dinosaurs eating pineapples.

Again on dx10 games tessellation is not being used dx11 path is not being used but clearly 5870 is getting beat by 4870x2 with lower core and high bandwidth on dx10 games. You can blame the architecture and drivers but it's too far fetched when evidence is right before your eyes.

 

[Daedalus685]

Comparing E6600 to E8400 and RV770 to cypress is within reason not dinosaurs eating pineapples.

Again on dx10 games tessellation is not being used dx11 path is not being used but clearly 5870 is getting beat by 4870x2 with lower core and high bandwidth on dx10 games. You can blame the architecture and drivers but it's too far fetched when evidence is right before your eyes.

Don't go telling me to look at evidence when you can't produce any, and refuse to see anything we psot about bandwidth scaling.

Ignoring the 4870x2 bandwidth confusion you have for a moment, yes, the 4870x2 does more than likely have a tad bit more bandwidth than a 5870, but not much.

No, tesselation, and much of the hardware for dx11 is not "used" during dx10.. but again, it is still THERE.

Why is it so far fetched that more hardware may add delays here and there, and effect the whole chip? The instructions still have to be sorted, dx or 11, it all goes through the same GPU, it is not as if there is a separate GPU core for dx11. Much of the new hardware is shared, it is perfectly rational to think this might change things.

Drivers are always a problem when a new chip comes out. Which is why many think this is the issue now too.

The only hard data we have is that adding more bandwidth does not effect the performance in a way that it would in a memory staved card. We have no evidence that the 5870 specific hardware is not affecting things, and we have no evidence that it is or is not drivers. Based on history we can assume the drivers are not perfect, and performance will improve.

We can not use the 4000 series as a comparison as they are different, you can't change more than the single variable you are trying to analyze.. using the 4870 changes all three.. It is unscientific nonsense.

What is the problem with that line of reasoning?

This is not a CPU die shrink.. Comparing the 4870 to the 4770 would be akin to the E8200 and E6600, this is not the same thing at all. Despite your assertions that it is. Just saying it does not make it so.

 

[Schmide]

Did anyone catch the cache? Twice? The 5000 series has an additional separate SIMD 8k L1 (16k vs 24k) cache for and double the L2 (64k vs 128k). No bandwidth can make up for what you don't have to fetch.

By processor analogy, the core architecture had less bandwidth than the Athlon 64 series, yet no one would argue advantage there.

 

[AzN]

Don't go telling me to look at evidence when you can't produce any, and refuse to see anything we psot about bandwidth scaling.

Quite contrary. There are multiple examples of bandwidth prevailing in this same thread. it seems you don't want to believe these examples but some raid array to make examples of crossfire. Blame drivers, blame architecture, blame me... bla bla..

Ignoring the 4870x2 bandwidth confusion you have for a moment, yes, the 4870x2 does more than likely have a tad bit more bandwidth than a 5870, but not much.

With no evidence of your claims that 4870x2 only has slight more bandwidth I might add but look below with my evidence that 4870x2 has more bandwidth than you seem to believe. If what you say is true wouldn't 5870 outperform 4870x2 since it's clocked much higher? But of course not. 4870x2 is the faster card with more bandwidth. What next? Blame drivers and architecture?

Testing high color fill is a great example of bandwidth limitation. Notice 4870x2 is 50% or so faster than 5870 in this test with less core clocks.

No, tesselation, and much of the hardware for dx11 is not "used" during dx10.. but again, it is still THERE.

Which 4870x2 has tessellation as well but you don't see me calling foul that it's just sitting there.

Why is it so far fetched that more hardware may add delays here and there, and effect the whole chip? The instructions still have to be sorted, dx or 11, it all goes through the same GPU, it is not as if there is a separate GPU core for dx11. Much of the new hardware is shared, it is perfectly rational to think this might change things.

More hardware like? tessellation again? dx11 path? Dx11 are nothing more than instructions and not hardware. We both know there's tessellation units in both rv770 and cypress.

Drivers are always a problem when a new chip comes out. Which is why many think this is the issue now too.

Did G92 have problems with drivers? Of course not... Most of the optimization is passed over from g80 and few numbers changed. it's usually when complete different architecture is where card manufacturers need some time to improve drivers.

The only hard data we have is that adding more bandwidth does not effect the performance in a way that it would in a memory staved card. We have no evidence that the 5870 specific hardware is not affecting things, and we have no evidence that it is or is not drivers. Based on history we can assume the drivers are not perfect, and performance will improve.

First off I never said 5870 was memory starved. You seem to assume I do. What I did say was that memory effects minimum fps and even slightly on avg frame rates where 5870 would be able to outperform 4870x2 with more bandwidth.

There's no evidence but we can test some of these scenarios by adding more bandwidth with 5850 with same core clocks as 5870 with higher memory clocks to show some results of memory limitations which there are examples in this same thread. There's nothing I can do for you not believing these examples and blame everything on drivers when there's no evidence.

We can not use the 4000 series as a comparison as they are different, you can't change more than the single variable you are trying to analyze.. using the 4870 changes all three.. It is unscientific nonsense.

You seem to frequent that word a lot. Unscientific nonsense? So testing 4000 series that is 97% of the same architecture is not scientific how cards behave to bandwidth? I don't quite agree with that assessment but whatever.

What is the problem with that line of reasoning?

This is not a CPU die shrink.. Comparing the 4870 to the 4770 would be akin to the E8200 and E6600, this is not the same thing at all. Despite your assertions that it is. Just saying it does not make it so.

But it's a GPU die shrink which corresponds with CPU die shrink.

SSE4 are instructions
dx11 are instructions

RV770 is 55nm
cypress is 40nm

 

[ugaboga232]

Azn we have given an entire 7 page thread full of info that thoroughly concludes that memory bandwidth is not the problem. A 50% increase in bandwidth gave a 12% increase in average frames. Before you start talking about minimum frames, very few sites give the data you want. You want a chart showing the fps by time and not just the lowest frame rate it reached.

I will say that before the research was done, yes memory bandwidth was the likeliest reason for the lower than expected performance. However with that being ruled out I think the following is the reason:

1. Drivers
2. AF implementation
3. Shaders not being utilized
4. Possibly hardware problems

Also AZN you are ignoring what I said before. The 5970 has the same problem as the 5870 aka being slower than expected but it has double the bandwidth. Why does that happen? If you continue to ignore me then you obviously do not know what you are talking about.

I think that this bottleneck either has been solved by AMD and will be fixed when the new Nvidia's come out or it is a hardware problem (unlikely imo).'

 

[AzN]

Azn we have given an entire 7 page thread full of info that thoroughly concludes that memory bandwidth is not the problem. A 50% increase in bandwidth gave a 12% increase in average frames. Before you start talking about minimum frames, very few sites give the data you want. You want a chart showing the fps by time and not just the lowest frame rate it reached.

You seem to be confused. What information did you exactly supply? Where is this 50% increase in bandwidth that gave 12% average frame rates? if you are talking about 4870x2 vs 5870 or 5770 vs 4890 it's more like 40% increase in bandwidth. I feel like a broken record. All I've been saying is that minimum frame rates is where bandwidth makes biggest difference and not so much on average. 5870 with same bandwidth as 4870x2 can easily beat the card. Never mind 4870x2 has bandwidth advantages. It's not drivers. simple as that.

I will say that before the research was done, yes memory bandwidth was the likeliest reason for the lower than expected performance. However with that being ruled out I think the following is the reason:

What research was done? The firingsquad overclocking bench? I simply see avg. frame rates for him to make that conclusion much like BFG 5770 testings. Forget minimum fps matter to most of us than average frame rates. It's an absurd conclusion without seeing more angles to the problem.

1. Drivers
2. AF implementation
3. Shaders not being utilized
4. Possibly hardware problems

Did you ever think that all the performance is already available with 5870. Without more clocking there is no performance to be had. I'm just saying with more bandwidth 5870 could be faster than 4870x2. It's simple concept you even posted above. You say 50% more bandwidth gives you 12% better frame rates. I don't exactly know for what card but let's say that applied to 5870 vs 4870x2 that is only 10% faster.

Also AZN you are ignoring what I said before. The 5970 has the same problem as the 5870 aka being slower than expected but it has double the bandwidth. Why does that happen? If you continue to ignore me then you obviously do not know what you are talking about.

I think that this bottleneck either has been solved by AMD and will be fixed when the new Nvidia's come out or it is a hardware problem (unlikely imo).'

What problem would this be? I don't see any problems but 1 fast card. Bottleneck is obvious bandwidth compared to cards before it. Blame drivers but when you pit a 5870 against 4870x2 you might want to place your bets on a card with 80% more bandwidth vs 12% reduced frequency.

 

[Daedalus685]

I do not, and have never said that memory BW could nto be the issue.. My oppinions are actually falsifiable unlike certain folks around here.

AZN, I simply believe, and have tried over and over again to explain, that it is very much possible that other issues coudl be at play.. I am not saying you are wrong.. just a bit silly with your blind stuborn nature.

I merely think that it is more likely due to drivers, but could also be hardware, maybe even BW.

As to why you are unscientific. You suffer from a severe confirmation bias. There are possible reasons that this behavior is seen that is not bandwidth but you ignore them. Thus you are comparing data sets that change these other possible variables becaseu you refuse to see them. If you were to do things properly you would only bring up data that changes one thing at a time, or at least as few as possible. Comparing a 4870x2 to a 5870 changes several of teh possibilities we have brought up. No matter how unlikely you think they are, it is still makes it unscientific to ignore the possibility by changing them all to try and prove a point. All you are acomplishing is showing that it could still be all of the reasons we have already brought up..

If you did this as your career you woudl ahve no job unless you learned to accept other possible answers. A good scientist always accepts that they could be wrong.. In fact science is about trying to prove you are wrong more so than proving you are right.. If you can't prove yourself wrong then you have a good theory. You do not have a good theory, and I repeat that it is nonsense, because all of the data you use as proof ignores the scientific method entirely.

If yuo want to compare things, choose one of drivers, bandwidth, and core architecture. Leave the others the same, and then compare. Proving that it coudl be memory BW does not prove you right, or us wrong. You do not seem to understand how one goes about dismissing a hypothesis.

For a hypothesis to be valid it there has to be a way to disprove it. To disprove something you have to prove it wrong.. but if there is no way to do this, "There is an elephant beside me," then the claim becomes nonsense. Repeatedly usign an example that does not prove you wrong is a sorry way to go about things...

 

[Daedalus685]

You seem to be confused. What information did you exactly supply? Where is this 50% increase in bandwidth that gave 12% average frame rates? if you are talking about 4870x2 vs 5870 or 5770 vs 4890 it's more like 40% increase in bandwidth. I feel like a broken record. All I've been saying is that minimum frame rates is where bandwidth makes biggest difference and not so much on average. 5870 with same bandwidth as 4870x2 can easily beat the card. Never mind 4870x2 has bandwidth advantages. It's not drivers. simple as that.

What research was done? The firingsquad overclocking bench? I simply see avg. frame rates for him to make that conclusion much like BFG 5770 testings. Forget minimum fps matter to most of us than average frame rates. It's an absurd conclusion without seeing more angles to the problem.

Did you ever think that all the performance is already available with 5870. Without more clocking there is no performance to be had. I'm just saying with more bandwidth 5870 could be faster than 4870x2. It's simple concept you even posted above. You say 50% more bandwidth gives you 12% better frame rates. I don't exactly know for what card but let's say that applied to 5870 vs 4870x2 that is only 10% faster.

What problem would this be? I don't see any problems but 1 fast card. Bottleneck is obvious bandwidth compared to cards before it. Blame drivers but when you pit a card against another with lower bandwidth you might want to place your bets on a card with 40% more bandwidth and 12% reduced frequency.

You really have to understand what minimum frame rate represents before you can understand anyone's point here. Any review lists the instantaneous minimum (im sick of having to tell you this). It does not represent anything. You want the sustained minimum for a time sample, or perhaps the integral. As it stands the average tells much more information than teh minimum as an instantaneous value is far too easy to be in error.

And again, comparing a 4870x2 to a 5870 is ignoring too many variables for it to mean anything.

 

[AzN]

AZN, I simply believe, and have tried over and over again to explain, that it is very much possible that other issues coudl be at play.. I am not saying you are wrong.. just a bit silly with your blind stuborn nature.

Quite simply I see your mass as ignorant blind nature. Quick to judge, discredit, blame, hate, follow the leader cause you think he cool, bla bla

Forget there's proof and examples. Forget I was right all along except call out on few people.

We might see things differently but did you ever think for a second that I saw little bit more than what you saw?

 

[Meaker10]

This is a pretty meaningless graph, it lets you get away with things that in games you would not be able to do.

Its a very crude way of getting the raw memory bandwidth. It would be like setting up the HD series with MAD operations and claiming they get true terraflop performance.

 

[Meaker10]

Also I would like to say that I have a card thats memory bandwidth limited.

A mobile HD 4650 DDR2, and let me tell you, clocking the memory gives LINEAR performance increases. If you are memory bandwidth limited it shows BIG TIME.

The HD 4670, clocked the same with GDDR3 clocked higher, gets that much higher results in 3dmark.

It does not put little drops here and there.