9800 GTX+/GTS 250

[BFG10K]

Originally posted by: Azn

I wouldn't care if BFG did the benchmarks. This wouldn't be the first time where he does the benchmarks only to find out I was right all along. His ultra core/sp/bandwidth reduction tests were done because of our arguments. He said SP makes the most difference while I said it was fillrate and right combination of bandwidth would win over SP long as the game wasn't SP limited back in late 2007.

Wow, some revisionist history from Anz there.

You first started this off by running around and claiming memory bandwidth was the primary factor. Then I posted benchmarks which proved you wrong, so you started running around claiming ?raw texture fillrate? (LMAO) was the primary limitation.

You were still doing this even after I linked to multiple benchmarks with AA that didn?t use FP texturing, thereby showing the increased texturing wasn?t even a factor in the cases you were claiming it was.

Now in this thread you appear to have ?forgotten? about texture fillrate and have changed your tune yet again, arguing something different (it looks like bandwidth again, but combined with shader performance).

You?ve changed your tune so many times I?m not even sure you even know what you?re arguing. Maybe we should take a lottery as to what you?ll come up with next year.

I was right about memory bandwidth and while the shader didn?t show as much difference as I thought on the 8800 Ultra, it most certainly showed a bigger relative difference on the GTX260+.

But since you think you know all of the cards I?ve owned and what tests I?ve run (when you don?t), you wouldn?t know any of this.

He only cites a card and then think all cards behave this way when in fact this isn't the case.

You mean like how you cite your sole and incomplete Crysis benchmark, and make sweeping generalizations on the G92 core when multiple independent results don?t agree with you?

I've been at this computer science thing for a long time if I sound pompous because I act like I know a lot about it is because I do. If that makes me pompous in your eyes so be it. I have no love for haters.

Heh, comedy gold, thanks for the good laugh.

 

[AzN]

Originally posted by: BFG10K
Again all you?re doing is simply confirming what I?ve been saying all along: that despite having reduced memory bandwidth, the 8800 GT is faster because of improvements to the core, thereby proving memory bandwidth isn?t the primary limiting factor.

You pretty much leech onto 1 thing. Bandwidth isn't the primary limiting factor when I'm explaining how that bandwidth limits a card like 8800gt when it can get much higher performance with more bandwidth. Card like 8800gts 640mb can't perform higher with more bandwidth much like 2900xt, 3870, 4870 and diminishing returns on the new GTX series but a g92 in the other hand have greater returns with more bandwidth. That's the point you don't quite seem to get.

Uh, no. My entire argument comes from the fact that you?re claiming memory is the limitation, yet parts still get faster when it?s reduced. Clearly it isn?t the primary limitation.

It?s like narrowing the neck of a bottle but still observing a higher flow of water. Clearly then the neck wasn?t primarily holding things back.

Your entire argument is based that memory bandwidth can be lowered and still perform faster but that's not even what this thread is about. We are trying to figure out if g92 is hindered by memory bandwidth. You must get it now. You go off based into different subject because that's the only point you have. Point taken. You can still get faster performance even if you lower bandwidth. non the less g92 is hindered by bandwidth.

Where to even start with this?

How about the fact that you used a single game, a single benchmark, and a single card and are trying to claim that sole result is somehow the norm when multiple G92 benchmarks disagree with you?

How about the fact that you never even touched the shader clock?

How about the fact that you didn?t even paint any kind of average across a range of scenarios?

Additionally, I?m pretty sure Chizow posted several tests from his G92 that showed the opposite of yours (i.e. memory showed the lowest performance impact). There were even several G92 overclocking threads that showed the same thing on the whole.

So yeah, your figures may be accurate, but they?re also an outlier. You absolutely cannot infer it?s the norm based on your sole result.

What did you use to determine G92 cards aren't bottlenecked by bandwidth? Oh that's right you used an ultra even cited a GTX275 that has nothing to do with g92.

Additionally you aren't sure Chizow posted several tests from his G92 because he doesn't have one. He used his GTX280 to show same type of results like the ultra or GTX275. I'm not arguing GTX and Ultra aren't core limited. I never said it wasn't but G92 has plenty of core and not enough bandwidth.

Again you seem to have a hard time understanding that if you reduce something but performance goes up, it?s not a limitation.

If everything was black and white that would be the case but it's not. You point to a card that has all the bandwidth the card needs with a weaker core and then compare it to a card that 35% more processing power, 3x texture fillrate with 10% lower bandwidth and then say it's not limited by bandwidth. You must see the hypocrisy in this but then again...

Again we know it isn?t because G92 parts beat other parts that have more bandwidth.

We also know it isn?t because G92 parts? relative performance increased despite having their bandwidth reduced.

That and you?ve repeatedly agreed that core improvements to the G92 more than offset the reduction in bandwidth because that?s the explanation you gave us as to why it?s faster.

Again you can compare to a card like 8600gt with 500gb/s of bandwidth and then compare to a card that has 5x of everything with 64gb/s of bandwidth and say it's not hindered by bandwidth. You must know for yourself how ridiculous this sounds but then again...

Oh, so you use the term ?limited? to describe anything that holds back performance? In that case every DX10 part is basically limited by everything (core/memory/shader/texturing/CPU) since it?s possible to find a scenario where improving one of these facets improves performance to some degree.

Heck, even overclocking the 2900XT?s VRAM can yield some performance improvement in certain situations. I guess using your terminology I can say the 2900XT is limited by bandwidth since moving the memory slider improves performance?

Using the term like that makes it lose meaning. More accurate usage would be based on the lowest proportionate ratio between the slider increase and the actual improvement. In that case the 4850 is clearly limited by the core, not the memory.

You must understand english? Limit, limited, limitation has the same meaning. Used differently to form a sentence. Since fillrate is handled through bandwidth only bandwidth can limit that fillrate from ever achieving its full potential. Which I've been explaining to you since 2007.

That?s my point, namely that the primary limitation is from the core.

Then again we aren't talking about RV770 that is pretty much balanced. G92 in the other hand is hindered by bandwidth and can separately clock shader, core, and memory to determine this which I already have.

But again these ?examples? have little to no basis in the real world. I?m showing you actual game figures while you post 3DMark synthetic tests as ?proof?. Uh-huh.

With your ultra and GTX? Again you haven't shown me anything relevant to g92.

Oh-oh Techreport used 3dmark synthetic tests as proof so did your partner. :laugh:

Uh, no, I don?t need to ?do? anything. Go assign homework exercises to someone who gives a shit.

If you strongly disagree with 3dmark but it's okay if he and techreport used 3dmark? But then bash someone else who cites techreport using 3dmark? I'm speechless. :Q

Is that supposed to incite some kind of a reaction on my part? Try harder.

Why don't you tell me why 2900xt beat 8800gtx in 3dmark2k6 instead? You must have a conclusion of some sort.

Again multiple independent benchmarks disagree with your sole test. If the card was hindered by memory bandwidth then it couldn?t possibly be showing performance gains by reducing it.

Multiple independent benchmarks that has nothing do with g92. point taken.

Yes, in your sole benchmark, which is neither complete or the norm.

Non the less it's a g92 testing a game. Never said it was the norm but then again g92 isn't the norm that's hindered in a big way by bandwidth.

Right, but again I?m not trying to imply bandwidth makes zero difference. Not to mention that the Gainward is sometimes faster than the 4870 despite having less bandwidth, clearly indicating driver or benchmarking noise.

So what you are saying 4850 is limited to bandwidth.

From my second post I stated no it?s not, not when it?s readily demonstrated that SP clocks generally have a bigger impact on performance than memory clocks. Again this is something that I?ve tested repeatedly with several parts.

Again my point is that bandwidth is not the primary limiting factor on DX10 parts (including the G92), not that bandwidth makes zero difference.

But based on my first comment in this thread I can see I used a poor choice of words to convey this meaning.

Of course you used poor choice of words. if you said something like bandwidth isn't the primary limiting factor on most dx10 parts I would have agreed with you but in case of g92 it is limited by bandwidth because it has all this core but no bandwidth to use it efficiently compared to other dx10 parts.

 

[AzN]

Originally posted by: BFG10K

Originally posted by: Azn

I wouldn't care if BFG did the benchmarks. This wouldn't be the first time where he does the benchmarks only to find out I was right all along. His ultra core/sp/bandwidth reduction tests were done because of our arguments. He said SP makes the most difference while I said it was fillrate and right combination of bandwidth would win over SP long as the game wasn't SP limited back in late 2007.

Wow, some revisionist history from Anz there.

You first started this off by running around and claiming memory bandwidth was the primary factor. Then I posted benchmarks which proved you wrong, so you started running around claiming ?raw texture fillrate? (LMAO) was the primary limitation.

You were still doing this even after I linked to multiple benchmarks with AA that didn?t use FP texturing, thereby showing the increased texturing wasn?t even a factor in the cases you were claiming it was.

Now in this thread you appear to have ?forgotten? about texture fillrate and have changed your tune yet again, arguing something different (it looks like bandwidth again, but combined with shader performance).

You?ve changed your tune so many times I?m not even sure you even know what you?re arguing. Maybe we should take a lottery as to what you?ll come up with next year.

I was right about memory bandwidth and while the shader didn?t show as much difference as I thought on the 8800 Ultra, it most certainly showed a bigger relative difference on the GTX260+.

But since you think you know all of the cards I?ve owned and what tests I?ve run (when you don?t), you wouldn?t know any of this.

You can tell me what you like but we still have evidence from 2007.

This is what you said....

Not that much given most games these days are shader bound. Witness how well the 8800 GT and 3870 do with just a 256 bit memory bus compared to 320/384/512 derivatives.

This is what I said...

I wouldn't say just shader bound with current generation of games. I would say it's more texture bound and feeding right amount of memory bandwidth than anything else. 8800gt does so well in this dept and pull ahead of original 8800gts because of this. If the new 8800gt had 384bit memory bus like gtx it would stomp an ultra in it's tracks.

Then you came around and said this...

Most of the bottleneck these days is with shaders given each pixel can have hundreds if not thousands instructions running on it before it?s drawn. The main reason it pulls away from the GTS is because it has more stream processors (aka shaders), which is exactly what I was saying earlier. Texturing is certainly important but memory bandwidth isn't that important unless you're really crippled on 128 bit or something. I doubt that very much; witness how the 2900 XT with ~50% more memory bandwidth isn't really faster than the 3870 because the rest of the specs are pretty much the same. Clearly memory bandwidth isn't having much of an impact in that situation. Similarly the Ultra has ~80% more memory bandwidth than the GT but it's not even close to being 80% faster so again memory bandwidth isn't the limiting factor there.

Few months later you tested your ultra only to find out core followed by bandwidth made the biggest difference.


Clearly your memory has failed you.

http://forums.anandtech.com/me...126198&highlight_key=y

Don't give me that shit about GTX series performance are swayed more by shader because Derek's exact benches show other wise. Biggest performance came from core and feeding right amount of memory bandwidth over shader even though there are more shader bound games today.

You mean like how you cite your sole and incomplete Crysis benchmark, and make sweeping generalizations on the G92 core when multiple independent results don?t agree with you?

Then again my card was a g92 and a real game benchmark. You on the other hand cite cards that has nothing to do with g92 and been saturated by bandwidth to death.

Heh, comedy gold, thanks for the good laugh.

Keys is comedy. I thought he needed it. :laugh:

 

[AzN]

Originally posted by: Scali
On the contrary of what?
Not sure how what you say relates to what I say.

Maybe we are in different waves.

I think you are looking at it from the opposite direction. What I'm saying is that it's G92 that is 'still just as fast' as G80 despite having a narrower memory interface and less ROPs. So G92 has a more efficient balance than G80 did. Because of the 256-bit bus they could use a cheaper PCB, and they used less memory to keep the cost down, without sacrificing performance. Quite a feat, really.

It looks we are looking at opposite direction. I find G80 to be more balanced card than G92. G92 can get better performance while G80 have diminishing returns with more bandwidth.

You'll also see that 8800GTX/Ultra is generally faster when you go for really high resolutions and high AA settings. The combination of the extra memory, the extra bandwidth and the extra fillrate will then kick in.

Exactly. In lower resolutions and without AA G92 usually beats the ultra. if g92 has more bandwidth it would stomp the ultra in high resolutions and aa settings.

 

[Scali]

Originally posted by: Azn
It looks we are looking at opposite direction. I find G80 to be more balanced card than G92. G92 can get better performance while G80 have diminishing returns with more bandwidth.

Well, there's the thing, you're calling G80 and G92 cards, when they're really just codenames for the GPU, and multiple variations of cards with these chips exist, with different balances between processing power and memory bandwidth.
I was just considering the GPU by itself, and the fact that G92 can get more performance out of less bandwidth than the G80. That's why I call it more balanced.

Indeed, when you clock the G92 high enough, you run into the problem that GDDR3 can't really go beyond 2.2 GHz. That's where your bandwidth bottleneck comes from.
You either need GDDR5, or you need a wider bus.
The problem with G80 is the opposite. You can get more bandwidth to the chip, but it can't do as much with that bandwidth.

Who knows though, we might see a G92-derivative on 40 nm with GDDR5, if the new line of mobile chips is any indication. That would solve the bandwidth problem, and we'll get to see what the true potential of the G92 architecture is.

 

[AzN]

Originally posted by: Scali

Originally posted by: Azn
It looks we are looking at opposite direction. I find G80 to be more balanced card than G92. G92 can get better performance while G80 have diminishing returns with more bandwidth.

Well, there's the thing, you're calling G80 and G92 cards, when they're really just codenames for the GPU, and multiple variations of cards with these chips exist, with different balances between processing power and memory bandwidth.
I was just considering the GPU by itself, and the fact that G92 can get more performance out of less bandwidth than the G80. That's why I call it more balanced.

You and I have a different way of looking at what you would call a balanced card. When I say it's more balanced I'm pointing at the fillrate that comprehend well with the available bandwidth and the fillrate not being wasted.

Indeed, when you clock the G92 high enough, you run into the problem that GDDR3 can't really go beyond 2.2 GHz. That's where your bandwidth bottleneck comes from.
You either need GDDR5, or you need a wider bus.
The problem with G80 is the opposite. You can get more bandwidth to the chip, but it can't do as much with that bandwidth.

Who knows though, we might see a G92-derivative on 40 nm with GDDR5, if the new line of mobile chips is any indication. That would solve the bandwidth problem, and we'll get to see what the true potential of the G92 architecture is.

My point exactly.

 

[Scali]

Originally posted by: Azn
You and I have a different way of looking at what you would call a balanced card. When I say it's more balanced I'm pointing at the fillrate that comprehend well with the available bandwidth and the fillrate not being wasted.

As I said, I'm talking about the GPU, not the card.
Since a GPU doesn't have memory, you can't include the available bandwidth into consideration.
The bandwidth depends on what you put on the card, not on the GPU itself.

 

[BFG10K]

Originally posted by: Azn

You pretty much leech onto 1 thing. Bandwidth isn't the primary limiting factor when I'm explaining how that bandwidth limits a card like 8800gt when it can get much higher performance with more bandwidth.

I?m not disputing it can show a performance gain by adding more bandwidth; my point is that bandwidth isn?t the primary limitation, even on the G92. Again we can repeatedly see this with the 8800 GT and 8800 GTS 512 beating cards like the 8800 GTS 640 while having less bandwidth. Clearly the primary performance gain is coming from somewhere other than from bandwidth, and the G92?s bandwidth is not crippled enough to offset that gain.

Card like 8800gts 640mb can't perform higher with more bandwidth much like 2900xt, 3870, 4870 and diminishing returns on the new GTX series but a g92 in the other hand have greater returns with more bandwidth. That's the point you don't quite seem to get.

See above. I can accept the G92 is more bandwidth bound than other DX10 parts (it must be given it has less bandwidth), but I don?t accept that it?s primarily bound by bandwidth like you claim.

Your entire argument is based that memory bandwidth can be lowered and still perform faster but that's not even what this thread is about. We are trying to figure out if g92 is hindered by memory bandwidth. You must get it now. You go off based into different subject because that's the only point you have. Point taken. You can still get faster performance even if you lower bandwidth. non the less g92 is hindered by bandwidth.

Okay, that?s fine, but again you can say any part is hindered by shader/core/memory/CPU/texturing since increasing any of them can show a some kind of performance gain in certain situations. What I?m talking about out is where the biggest hindrance is.

What did you use to determine G92 cards aren't bottlenecked by bandwidth?

By checking the benchmark charts and confirming they aren?t stacked relative to memory bandwidth, but according to certain elements of their core.

Additionally you aren't sure Chizow posted several tests from his G92 because he doesn't have one.

I?m sure now:

http://forums.anandtech.com/me...AR_FORUMVIEWTMP=Linear

GT @650/850 (Stock 8800GT SC)
2007-11-20 11:27:16 - lotroclient
Frames: 7174 - Time: 120000ms - Avg: 59.783 - Min: 30 - Max: 83

GT @650/1000
2007-11-20 11:32:31 - lotroclient
Frames: 7294 - Time: 120000ms - Avg: 60.783 - Min: 31 - Max: 85

GT @675/1000
2007-11-20 11:36:28 - lotroclient
Frames: 7437 - Time: 120000ms - Avg: 61.975 - Min: 33 - Max: 87

GT @700/1000
2007-11-20 11:41:17 - lotroclient
Frames: 7467 - Time: 120000ms - Avg: 62.225 - Min: 25 - Max: 96

GT @729/1000
2007-11-20 11:50:40 - lotroclient
Frames: 7611 - Time: 120000ms - Avg: 63.425 - Min: 33 - Max: 105

GT @729/1050 (Unstable in ATITool)
2007-11-20 11:59:53 - lotroclient
Frames: 7601 - Time: 120000ms - Avg: 63.342 - Min: 28 - Max: 102

An 18% memory overclock improved performance by 2%, but it took only a 4% overclock on the core to achieve the same performance gain. It?s painfully obvious that the bulk of the performance gain is coming from raising the core, not from the memory.

I never said it wasn't but G92 has plenty of core and not enough bandwidth.

Again if it had ?plenty of core? we wouldn?t be seeing bigger performance improvements from core/shader clock increases than from memory, but we are. So even with the reduced bandwidth the core/shader is still the primary bottleneck on a G92.

If everything was black and white that would be the case but it's not. You point to a card that has all the bandwidth the card needs with a weaker core and then compare it to a card that 35% more processing power, 3x texture fillrate with 10% lower bandwidth and then say it's not limited by bandwidth. You must see the hypocrisy in this but then again...

What hypocrisy? It?s a basic observation how the improved core is still able to improve performance overall despite having less bandwidth to use.

Again you can compare to a card like 8600gt with 500gb/s of bandwidth and then compare to a card that has 5x of everything with 64gb/s of bandwidth and say it's not hindered by bandwidth. You must know for yourself how ridiculous this sounds but then again...

If the 64 GB/sec card is still showing a greater performance gain from shader/core than memory then I can absolutely state it?s not primarily limited by bandwidth. That?s my whole point.

Since fillrate is handled through bandwidth only bandwidth can limit that fillrate from ever achieving its full potential. Which I've been explaining to you since 2007.

What utter nonsense. Shader performance can limit fillrate because if the ROPs aren?t fed fast enough, fillrate is squandered because ROPs sit around waiting for data to process. So can triangle setup. So can vertex processing. So can anything else that comes in the pipeline before the ROPs.

To claim bandwidth is the only thing that can limit fillrate is comically nonsensical.

Now, perhaps a synthetic fillrate benchmark like 3DMark is immune to this, but this simply shows how useless such a synthetic benchmark is given it can?t show the full picture.

G92 in the other hand is hindered by bandwidth and can separately clock shader, core, and memory to determine this which I already have.

Like I said, there?s a plethora of examples that prove your example is an outlier.

Oh-oh Techreport used 3dmark synthetic tests as proof so did your partner. :laugh:

Good for them. And?

If you strongly disagree with 3dmark but it's okay if he and techreport used 3dmark? But then bash someone else who cites techreport using 3dmark?

I bash someone for citing 3DMark figures as evidence to an argument where they have absolutely no correlation to reality.

Why don't you tell me why 2900xt beat 8800gtx in 3dmark2k6 instead? You must have a conclusion of some sort.

Yes, I have a conclusion, but the only conclusion that really matters is that 3DMark is a useless synthetic test that tells you nothing about how the cards will behave in real games.

 

[Scali]

Originally posted by: BFG10K
I?m not disputing it can show a performance gain by adding more bandwidth; my point is that bandwidth isn?t the primary limitation, even on the G92. Again we can repeatedly see this with the 8800 GT and 8800 GTS 512 beating cards like the 8800 GTS 640 while having less bandwidth. Clearly the primary performance gain is coming from somewhere other than from bandwidth, and the G92?s bandwidth is not crippled enough to offset that gain.

I think there are two different things going on here:
1) G92 makes use of bandwidth more efficiently than G80, hence more performance per unit of bandwidth.
2) G92 may or may not be limited by bandwidth.

Because the efficiency of G80 and G92 aren't equal, you can't do 1:1 comparisons.
Even if G92 is faster with the same or less bandwidth than G80 is, that is no guarantee that G92 isn't bandwidth-limited. It could still be bandwidth-limited. G80 will just run into a bandwidth-limit at a lower level of performance, since it doesn't use its bandwidth as efficiently.

To put it another way... Say you have a G92 card and a G80 card, both with the exact same amount of bandwidth. You run a benchmark. The G92 card gets 50 fps, the G80 card gets 45 fps.
In that scenario they could still BOTH be bandwidth-limited. You can't really conclude that from these numbers. Even though G92 squeezes more frames out of the same bandwidth, that doesn't mean that the bandwidth isn't the limiting factor.

To give a similar example: Even though Core2 systems and Phenom systems both use dualchannel DDR2 or DDR3 memory, the Phenom systems get more effective bandwidth and lower latency out of the exact same memory, because they use an onboard memory controller, which is more efficient.
It's the same theoretic bandwidth from the memory, but in a complete system, the results are different.

 

[BFG10K]

Originally posted by: Azn

This is what I said...

I wouldn't say just shader bound with current generation of games. I would say it's more texture bound and feeding right amount of memory bandwidth than anything else. 8800gt does so well in this dept and pull ahead of original 8800gts because of this. If the new 8800gt had 384bit memory bus like gtx it would stomp an ultra in it's tracks.

Right, and we all know how wrong you were about texturing. The 9600 GT for example has greatly reduced texturing performance, yet it?s basically the same speed as the 8800 GT.

Furthermore my shader comments were backed by multiple review sites along with trend calculations from ATi. Even your own beloved Tech-Report commented on shader performance becoming progressively more important than texturing.

Few months later you tested your ultra only to find out core followed by bandwidth made the biggest difference.

Not quite. The core made the biggest difference yes, but without AA the shader was easily second place, with memory being a distant third. Even after factoring in AA, the shader core was still slightly ahead of memory overall.

Given the core was about the same for both and that we expect AA to hit the memory harder, it proved without a doubt the 2006/2007 games we talked about still significantly benefitted from shader clocks.

Now it?s quite true, I was surprised the core made the biggest impact in both scenarios, but this could simply be because the shader performance on the 8800 Ultra relative to the core was in excess of what those older titles required.

Don't give me that shit about GTX series performance are swayed more by shader because Derek's exact benches show other wise. Biggest performance came from core and feeding right amount of memory bandwidth over shader even though there are more shader bound games today.

Again, you clearly have no idea what tests I?ve run and on what cards I?ve run them on. Furthermore the card you called ?bandwidth happy? (GTX260+) actually experienced a relatively even mix of performance changes when core/shader/memory clocks were adjusted, indicating it?s quite a balance part.

Then again my card was a g92 and a real game benchmark.

As were repeated tests I linked to.

 

[BFG10K]

Originally posted by: Scali

In that scenario they could still BOTH be bandwidth-limited. You can't really conclude that from these numbers. Even though G92 squeezes more frames out of the same bandwidth, that doesn't mean that the bandwidth isn't the limiting factor.

That certainly could be true, but you can still show the G92 isn?t primarily limited by bandwidth by demonstrating the core clock impacts performance more than the memory clock does.

Then you don?t even need to compare theoretical bandwidth between G80 and G92, you just compare the same card to itself.

 

[AzN]

Originally posted by: BFG10K
I?m not disputing it can show a performance gain by adding more bandwidth; my point is that bandwidth isn?t the primary limitation, even on the G92. Again we can repeatedly see this with the 8800 GT and 8800 GTS 512 beating cards like the 8800 GTS 640 while having less bandwidth. Clearly the primary performance gain is coming from somewhere other than from bandwidth, and the G92?s bandwidth is not crippled enough to offset that gain.

I'm not disputing bandwidth is the primary limiting factor. I never claimed it did. What I'm proposing is that bandwidth and fillrate goes hand in hand and effect the way gpu behave. This is the same reason why 2900xt is useless with 512bit memory controller and perform same as 3870 with 256bit memory controller. While some cards get bigger gains with more bandwidth like my 8800gts did.

See above. I can accept the G92 is more bandwidth bound than other DX10 parts (it must be given it has less bandwidth), but I don?t accept that it?s primarily bound by bandwidth like you claim.

Finally thank you. That's all I'm proposing. G92 is more bandwidth bound than other dx10 parts and can perform just as fast as GTX260 if it had more bandwidth. After all that is what the thread is all about...

Okay, that?s fine, but again you can say any part is hindered by shader/core/memory/CPU/texturing since increasing any of them can show a some kind of performance gain in certain situations. What I?m talking about out is where the biggest hindrance is.

When I mention hindered I'm talking about g92 being hindered by bandwidth as per title of this thread. Not what part of the GPU is biggest hindered as this would change for different GPU~different generation of games.

By checking the benchmark charts and confirming they aren?t stacked relative to memory bandwidth, but according to certain elements of their core.

This has nothing to do with G92. Why? Because all cards do not behave same way.

I?m sure now:

http://forums.anandtech.com/me...AR_FORUMVIEWTMP=Linear

An 18% memory overclock improved performance by 2%, but it took only a 4% overclock on the core to achieve the same performance gain. It?s painfully obvious that the bulk of the performance gain is coming from raising the core, not from the memory.

I never got around and replied to that post. I must have been away.

Chizow does not mention what resolution nor AA settings. He could be running 1280 with no AA. Considering a full g92 chip is only 10% slower on average than GTX260 with no AA and 30% slower with AA his benches seem like it wasn't done with any form of AA at all. That could be one of the reason why the memory bandwidth couldn't have effected his benchmarks more so. Incidentally he doesn't leave his memory clocks at stock and raise just the core clocks to determine this either. He determines this by using combination of core and memory clocks when he upped the core benches which doesn't paint a clearer picture. Never the less he's getting less performance by adding more bandwidth in his last benchmark of 729/1050 vs 729/1000. His benches seem all over the place looking at some of the results of minimum/maximum frame rates which could mean his benches are flawed.

I used Crysis because this game is GPU limited in a big way. Unfortunately I don't have my G92 8800gts anymore to give you more refined results.

8800gt also has 48TMU 112SP and has nearly the same bandwidth as my G92 8800gts which has 64TMU and 128SP. This could also reflect how bandwidth doesn't effect 8800gt as much as full g92.

Incidentally Archangel did some benches of his own in that same thread with AA which was effected in a dramatic way. Of course this is more of a extreme situation and I understand this. 4xAA should be the standard not 16xAA.

Originally posted by: ArchAngel777
8800GTS 512 MB.

FEAR @ 1680 X 1050 16XAA/16XAF, TSAA, HQ, SOFT SHADOWS ON EVERYTHING ON MAX

750/1825/1600

16 Min
30 Avg

750/1825/2200

20 Min
39 Avg

Again if it had ?plenty of core? we wouldn?t be seeing bigger performance improvements from core/shader clock increases than from memory, but we are. So even with the reduced bandwidth the core/shader is still the primary bottleneck on a G92.

Again my explanation above. Chizow is probably using no AA and using relatively low resolution which reflect those kind of results. 8800gt is only 80% of the full g92 which I used a full g92 to determine this and not some cut version of the core.

What hypocrisy? It?s a basic observation how the improved core is still able to improve performance overall despite having less bandwidth to use.

Then again my benches show other wise. that's why we need more extensive testing on g92 on modern games. Before your pointing chizow benches you've only pointed to cards that have plenty of bandwidth like ultra or GTX275 and you trying to compare it to G92. Was that even fair or have any relevance?

If the 64 GB/sec card is still showing a greater performance gain from shader/core than memory then I can absolutely state it?s not primarily limited by bandwidth. That?s my whole point.

Again I never implied bandwidth was the greatest limiting factor. I've repeated this multiple times. What I did propose was that bandwidth and fillrate work side by side and some cards are more limited to bandwidth than others.

What utter nonsense. Shader performance can limit fillrate because if the ROPs aren?t fed fast enough, fillrate is squandered because ROPs sit around waiting for data to process. So can triangle setup. So can vertex processing. So can anything else that comes in the pipeline before the ROPs.

To claim bandwidth is the only thing that can limit fillrate is comically nonsensical.

Now, perhaps a synthetic fillrate benchmark like 3DMark is immune to this, but this simply shows how useless such a synthetic benchmark is given it can?t show the full picture.

What BS. Shader performance does not limit fillrate. It has nothing to do with fillrate. You can go ahead and raise the clocks of your shader and test any type of benches that corresponds with fillrate which would show SP does not limit fillrate.

Like I said, there?s a plethora of examples that prove your example is an outlier.

Like your ultra? Anand's GTX275. Chizow's 8800gt with no AA and unknown resolution. How about Archangel's Fear with 16xAA? My full g92 benchmark is outlier but other benchmarks that has nothing to do with G92 is. How convenient!

I bash someone for citing 3DMark figures as evidence to an argument where they have absolutely no correlation to reality.

No you bash someone citing techreport who used 3dmark but not techreport for using 3dmark to get a conclusion.

Yes, I have a conclusion, but the only conclusion that really matters is that 3DMark is a useless synthetic test that tells you nothing about how the cards will behave in real games.

You still can't come up with a conclusion. Figures..

 

[BenSkywalker]

What BS. Shader performance does not limit fillrate. It has nothing to do with fillrate. You can go ahead and raise the clocks of your shader and test any type of benches that corresponds with fillrate which would show SP does not limit fillrate.

If anything is shader limited then shader performance will be limiting the fillrate always. You don't need to run tests to prove this, just a basic understanding of elementary factors.

You still can't come up with a conclusion. Figures..

It would be idiotic to come up with one conclusion. Different things are going to limit you at different times. Even within a given bench of say Crysis, the CPU is going to be the limiting factor at certain points even using very high settings. What you can try to do is figure out what is the element that is holding back performance the most often, given all evidence we have in current games that appears to be shader performance.

 

[AzN]

Originally posted by: BFG10K

Originally posted by: Azn

This is what I said...

I wouldn't say just shader bound with current generation of games. I would say it's more texture bound and feeding right amount of memory bandwidth than anything else. 8800gt does so well in this dept and pull ahead of original 8800gts because of this. If the new 8800gt had 384bit memory bus like gtx it would stomp an ultra in it's tracks.

Right, and we all know how wrong you were about texturing. The 9600 GT for example has greatly reduced texturing performance, yet it?s basically the same speed as the 8800 GT.

Furthermore my shader comments were backed by multiple review sites along with trend calculations from ATi. Even your own beloved Tech-Report commented on shader performance becoming progressively more important than texturing.

Nope not wrong about 9600gt. What you still don't get is that memory bandwidth is holding back the 8800gt. With AA 9600gt was close to 8800gt level of performance because it had same amount of bandwidth. Without AA 8800gt was 25-30% faster and sometimes faster than 9600gt. This was prevalent when we disabled shader with rivatuner when 9600gt was released and tested games with it.

Furthermore your shader comments weren't backed by multiple review sites because you never implied this before these sites mentioned them. My beloved tech-report commented shader performance becoming progressively more important doesn't mean it has.

Not quite. The core made the biggest difference yes, but without AA the shader was easily second place, with memory being a distant third. Even after factoring in AA, the shader core was still slightly ahead of memory overall.

Given the core was about the same for both and that we expect AA to hit the memory harder, it proved without a doubt the 2006/2007 games we talked about still significantly benefitted from shader clocks.

Now it?s quite true, I was surprised the core made the biggest impact in both scenarios, but this could simply be because the shader performance on the 8800 Ultra relative to the core was in excess of what those older titles required.

So you finally agree which I've been saying since 2007 that shader isn't the biggest limiting factor but combination of fillrates with bandwidth long as the game isn't shader limited.

But you see without bandwidth fillrate wouldn't even get anywhere it needs to go. If your ultra had 64bit memory controller that core would make less of a difference. that's the point I'm trying to make which I've saying in that 2007 thread as well.

And who the hell plays a game without AA these days? I remember people were praising 9600gt for having better AA performance than 8800gs although 8800gs raw frames were usually faster. Now you want to go back to citing games without AA like that has any relevance these days?

Again, you clearly have no idea what tests I?ve run and on what cards I?ve run them on. Furthermore the card you called ?bandwidth happy? (GTX260+) actually experienced a relatively even mix of performance changes when core/shader/memory clocks were adjusted, indicating it?s quite a balance part.

Well Derek's benches show other wise. If you have something to share that backs up your claims I suggest you share them instead of telling what if's.

As were repeated tests I linked to.

No you pointed to GTX275 and the ultra benches.

 

[AzN]

Originally posted by: BenSkywalker

What BS. Shader performance does not limit fillrate. It has nothing to do with fillrate. You can go ahead and raise the clocks of your shader and test any type of benches that corresponds with fillrate which would show SP does not limit fillrate.

If anything is shader limited then shader performance will be limiting the fillrate always. You don't need to run tests to prove this, just a basic understanding of elementary factors.

That would depend if the game was shader limited in the first place.

You still can't come up with a conclusion. Figures..

It would be idiotic to come up with one conclusion. Different things are going to limit you at different times. Even within a given bench of say Crysis, the CPU is going to be the limiting factor at certain points even using very high settings. What you can try to do is figure out what is the element that is holding back performance the most often, given all evidence we have in current games that appears to be shader performance.

[/quote]

I want to hear his conclusion instead of just bashing the benchmarking tool. Whatever conclusion he has is an educated guess. I wouldn't bash him for that. I just want to hear what he has to say.

 

[AzN]

Originally posted by: Scali

Originally posted by: BFG10K
I?m not disputing it can show a performance gain by adding more bandwidth; my point is that bandwidth isn?t the primary limitation, even on the G92. Again we can repeatedly see this with the 8800 GT and 8800 GTS 512 beating cards like the 8800 GTS 640 while having less bandwidth. Clearly the primary performance gain is coming from somewhere other than from bandwidth, and the G92?s bandwidth is not crippled enough to offset that gain.

I think there are two different things going on here:
1) G92 makes use of bandwidth more efficiently than G80, hence more performance per unit of bandwidth.
2) G92 may or may not be limited by bandwidth.

Because the efficiency of G80 and G92 aren't equal, you can't do 1:1 comparisons.
Even if G92 is faster with the same or less bandwidth than G80 is, that is no guarantee that G92 isn't bandwidth-limited. It could still be bandwidth-limited. G80 will just run into a bandwidth-limit at a lower level of performance, since it doesn't use its bandwidth as efficiently.

To put it another way... Say you have a G92 card and a G80 card, both with the exact same amount of bandwidth. You run a benchmark. The G92 card gets 50 fps, the G80 card gets 45 fps.
In that scenario they could still BOTH be bandwidth-limited. You can't really conclude that from these numbers. Even though G92 squeezes more frames out of the same bandwidth, that doesn't mean that the bandwidth isn't the limiting factor.

To give a similar example: Even though Core2 systems and Phenom systems both use dualchannel DDR2 or DDR3 memory, the Phenom systems get more effective bandwidth and lower latency out of the exact same memory, because they use an onboard memory controller, which is more efficient.
It's the same theoretic bandwidth from the memory, but in a complete system, the results are different.

My point exactly. Thank you for the explanation as I don't explain too well at times.

 

[BFG10K]

Originally posted by: Azn

I'm not disputing bandwidth is the primary limiting factor. I never claimed it did.

So to confirm then, you don?t think the G92 is primarily limited by bandwidth? If so then we agree, so we can move on from this part.

When I mention hindered I'm talking about g92 being hindered by bandwidth as per title of this thread. Not what part of the GPU is biggest hindered as this would change for different GPU~different generation of games.

Sure, but then my point is that a 8800 Ultra is hindered by bandwidth too given it shows a performance change with memory clocks.

This has nothing to do with G92. Why? Because all cards do not behave same way.

We can still compare a G92 cards to each other and see where they rank. We can still change the clocks on a single G92 card and see what changes.

Chizow does not mention what resolution nor AA settings.

Yes, he stated he was using 1920x1200. He?s been over this with you already and I?m not going to repeat what he said.

Furthermore, you were only running 1680x1050 weren?t you? That?s less than his resolution.

His benches seem all over the place looking at some of the results of minimum/maximum frame rates which could mean his benches are flawed.

LMAO, and there we have it. Any benchmark which doesn?t agree with your sole incomplete Crysis benchmark is either flawed, memory happy, or doesn?t count because 3DMark@Tech-Report says something else. :roll:

Then again my benches show other wise. that's why we need more extensive testing on g92 on modern games. Before your pointing chizow benches you've only pointed to cards that have plenty of bandwidth like ultra or GTX275 and you trying to compare it to G92. Was that even fair or have any relevance?

There?s plenty of evidence in any 8800 GT overclocking thread, several of which were linked to you in past discussions. Again, I?m not going though this again with you.

What BS. Shader performance does not limit fillrate. It has nothing to do with fillrate. You can go ahead and raise the clocks of your shader and test any type of benches that corresponds with fillrate which would show SP does not limit fillrate.

Now I see why you don?t get it, and why you don?t understand why synthetic tests are worthless.

If I have a pixel fillrate of 10 pixels/sec, but my shaders can only shade enough data for 8 pixels/sec when used, my fillrate will top out at 8 pixels/sec in those situations. So shaders can absolutely limit it. The maximum performance will always be limited by the slowest currently performing section in a state-machine like a 3D pipeline.

Since ROPs are last in the pipeline, everything from the CPU to PhysX to shaders ? and anything in between - can affect it.

Now, a synthetic benchmark like a fillrate test is designed only to test one part of a card, so it?ll be designed to touch as little else of the pipeline as possible. That?s is why synthetic benchmarks cannot paint a true picture of reality, and why it?s pointless to discuss 3DMark.

That would depend if the game was shader limited in the first place.

So you admit you were wrong when you claimed only memory bandwidth can limit fillrate?

Like your ultra? Anand's GTX275. Chizow's 8800gt with no AA and unknown resolution. How about Archangel's Fear with 16xAA? My full g92 benchmark is outlier but other benchmarks that has nothing to do with G92 is. How convenient!

No, like the plethora of 8800 GT overclocking figures online which have been linked to you multiple times in the past.

And like how the 9800GTX has 14% more memory bandwidth than a 8800 GTS 512 yet doesn?t come anywhere close to getting that kind of a speedup, once again proving memory bandwidth isn?t the primary factor on a G92.

It?s absolutely comical how you think your sole incomplete Crysis figure benchmark is the norm and then state other benchmarks are invalid.

No you bash someone citing techreport who used 3dmark but not techreport for using 3dmark to get a conclusion.

Tech-Report?s conclusion is irrelevant to this discussion because it?s based on synthetic tests.

You still can't come up with a conclusion. Figures..

Again, why waste time discussing something that clearly doesn?t reflect real-world performance?

 

[BFG10K]

Originally posted by: Azn

Nope not wrong about 9600gt. What you still don't get is that memory bandwidth is holding back the 8800gt. With AA 9600gt was close to 8800gt level of performance because it had same amount of bandwidth. Without AA 8800gt was 25-30% faster and sometimes faster than 9600gt. This was prevalent when we disabled shader with rivatuner when 9600gt was released and tested games with it.

http://www.behardware.com/arti...nd-radeon-hd-3650.html

At 1280x1024 with no AA (i.e. when memory bandwidth isn?t a factor), the 8800 GT is not even close to being faster proportionate to texturing differences. We?re seeing it being ~21% faster when the 8800 GT has 75% more texturing power than the 9600 GT.

Clearly then texturing performance isn?t the primary limitation like you claimed.

Furthermore at 1680x1050 with 4xAA, the performance gap increases where the 8800 GT is 42% faster than the 9600 GT. How can that be if you claim the 8800 GT is limited by bandwidth, but both cards have the same amount? And it can?t be texturing because texturing has nothing to do with AA.

Furthermore your shader comments weren't backed by multiple review sites because you never implied this before these sites mentioned them.

Right, but that?s the purpose of review sites, to learn about 3D technology. Since I doubt you?re an industry insider, you have to read them just like the rest of us mere mortals.

My beloved tech-report commented shader performance becoming progressively more important doesn't mean it has.

You mean kind of like of like how the 3DMark tests you quote aren?t relevant?

So you finally agree which I've been saying since 2007 that shader isn't the biggest limiting factor but combination of fillrates with bandwidth long as the game isn't shader limited.

I agree that in the case of the 8800 Ultra it had excess shader performance for those older titles. I?d love to retest newer titles but alas, the card is dead. But like I said, other cards I?ve tested (e.g. GTX260+) show a much more balanced change.

Again we?ve already been through this two years ago when I pointed out the performance gain of the X1900 vs the X1800 in current games by simply tripling the shader power.

And who the hell plays a game without AA these days? I remember people were praising 9600gt for having better AA performance than 8800gs although 8800gs raw frames were usually faster. Now you want to go back to citing games without AA like that has any relevance these days?

AA is very relevant, but for these kinds of discussions it also skews the results because it creates an external factor that drains bandwidth, so it doesn?t allow factors to come into light that would otherwise do so. So I?m saying it?s important to use both AA and no AA to get a clearer picture for this kind of discussion.

Well Derek's benches show other wise. If you have something to share that backs up your claims I suggest you share them instead of telling what if's.

Well sure, I guess I can link to this one since someone else linked to it first:

http://forums.anandtech.com/me...287281&highlight_key=y

No you pointed to GTX275 and the ultra benches.

The parts that say ?8800 GT? on the benchmark bars disagree with you.

 

[SickBeast]

Wow. I just pasted the last 100 posts in this thread in to MS Word and it came out to over 9,500 words. AZN and BFG10K, you two could have probably written 2-3 articles each instead of arguing endlessly in this thread. It didn't change anything, either. AZN has been trying to prove BFG wrong this entire time, while BFG has tried to prove himself right. I would suggest that AZN still does not believe him. :beer:

 

[AzN]

double post

 

[AzN]

Originally posted by: BFG10K

Originally posted by: Azn

Nope not wrong about 9600gt. What you still don't get is that memory bandwidth is holding back the 8800gt. With AA 9600gt was close to 8800gt level of performance because it had same amount of bandwidth. Without AA 8800gt was 25-30% faster and sometimes faster than 9600gt. This was prevalent when we disabled shader with rivatuner when 9600gt was released and tested games with it.

http://www.behardware.com/arti...nd-radeon-hd-3650.html

At 1280x1024 with no AA (i.e. when memory bandwidth isn?t a factor), the 8800 GT is not even close to being faster proportionate to texturing differences. We?re seeing it being ~21% faster when the 8800 GT has 75% more texturing power than the 9600 GT.

Clearly then texturing performance isn?t the primary limitation like you claimed.

That's because it's limited by bandwidth to some extent even without AA. Just because a card has 75% more texturing doesn't mean you get 75% better frame rates. This kind of linear thinking is why you've been wrong in 2007 which you finally admitted after 2 years of saying you were right.

Furthermore at 1680x1050 with 4xAA, the performance gap increases where the 8800 GT is 42% faster than the 9600 GT. How can that be if you claim the 8800 GT is limited by bandwidth, but both cards have the same amount? And it can?t be texturing because texturing has nothing to do with AA.

Again you only see black and then white. A game could have been SP hungry like Bioshock which hinder 9600gt performance @ 1680x1050 4xAA because it only has 64SP vs 112 on 8800gt.

Right, but that?s the purpose of review sites, to learn about 3D technology. Since I doubt you?re an industry insider, you have to read them just like the rest of us mere mortals.

I'm not an insider but what I'm really good at is analyzing things. I don't need some web site to tell me that shader limited games are on the rise. I can figure that out for myself when they do appear and make more difference in games.

You mean kind of like of like how the 3DMark tests you quote aren?t relevant?

Techreport tested and gave an analysis which I come to agree with when they did the pixel fill test because I've done these kind of testings as well. But what they aren't qualified is trying to foresee the future.

I agree that in the case of the 8800 Ultra it had excess shader performance for those older titles. I?d love to retest newer titles but alas, the card is dead. But like I said, other cards I?ve tested (e.g. GTX260+) show a much more balanced change. Again we?ve already been through this two years ago when I pointed out the performance gain of the X1900 vs the X1800 in current games by simply tripling the shader power.

Like I said. You finally agreed with what I've been saying all this time and that you were wrong after 2 years of BS you've been giving me the run around. Much like this thread. I'm not going to argue with you because it's becoming a big waste of time. I'll let your tests and my tests do the talking when you realize it for yourself in another 5 years!!!

 

[Scali]

Originally posted by: Azn
I'm not an insider but what I'm really good is at analyzing things. I don't need some web site to tell me that shader limited games are on the rise. I can figure that out for myself when they do appear and make more difference in games.

I don't want to get into who is right or wrong, but I must say it's refreshing that some people still like to think for themselves, and aren't afraid to put forward a view that may go against the mainstream.

In this day and age where people just google and read internet articles, things have become very boring. Most people just repeat what they read, regardless of whether or not they understand it. Even review sites can sometimes be wrong... in fact, they often are, because a review is only a 'snapshot' at a certain point in time. I think a great example would be when the first T&L hardware arrived. Most review sites ran standard games and benchmarks on them, and concluded that T&L acceleration was useless. It was just a gimmick, and you were better off with a fast CPU. What they didn't understand was that none of these games and benchmarks were designed to take advantage of T&L hardware, and as such were very inefficient. Later games ran much better on the same hardware, and can you even imagine running with software T&L these days?

Back in those days I used a Matrox G450, and upgraded to a GeForce2. I was totally blown away by the sheer amount of polygons I could suddenly stuff into the hardware with little or no performance hit. I could go from a few thousand polys per frame to hundreds of thousands, overnight.
Yet when I read articles and forum discussions regarding T&L, the majority was just bashing it as a gimmick and hailing the CPU as T&L king.
Just because the majority thinks otherwise doesn't mean they're right.

I would encourage more people to really think for themselves, conduct some proper analysis, and get a deeper understanding, so you can draw your own conclusions. You may not always be right, but being wrong is not such a big deal, is it? We all learn from our mistakes anyway. It would make for much more interesting discussion, rather than people just regurgitating what every other forum discussion and article says.

 

[BFG10K]

Originally posted by: Azn

That's because it's limited by bandwidth to some extent even without AA. Just because a card has 75% more texturing doesn't mean you get 75% better frame rates. This kind of linear thinking is why you've been wrong in 2007 which you finally admitted after 2 years of saying you were right.

If 1280x1024 is still bandwidth limited to make such a large difference, then why did you try to discount Chizow?s results on the basis that they might?ve been run at the same resolution? Is 1280x1024 bandwidth limited or not?

Also I?m not saying it has to be 75%, just more proportional to the increase relative to the other specs? changes. For an example, see here:

http://www.xbitlabs.com/articl...gt-512gs_12.html#sect1

Those are 1920x1200 results for UT3 and Stalker; note the large performance difference between the parts despite having identical memory bandwidth. In UT3 the 8800 GT is 67% faster than the 9600 GT. That?s in line with the differences to the core (shader and/or texturing) despite bandwidth being the same on both cards. There?s simply no way those 1920x1200 tests can be limited by memory with that kind of rift.

Like I said earlier, you won?t accept any result that is different to your sole Crysis test and that?s why you don?t see the full picture. That?s why it?s important to look at a range of benchmarks at a range of settings with this type of thing. In particular, averages are really important to stop erratic titles from skewing the results too far in one direction.

Again you only see black and then white. A game could have been SP hungry like Bioshock which hinder 9600gt performance @ 1680x1050 4xAA because it only has 64SP vs 112 on 8800gt.

Yes, it probably could be shaders but how does that factor into your memory bandwidth theories if it?s still shader bound at 4xAA? Clearly then memory bandwidth isn?t the primary bottleneck in that instance, even with 4xAA.

I'm not an insider but what I'm really good at is analyzing things. I don't need some web site to tell me that shader limited games are on the rise. I can figure that out for myself when they do appear and make more difference in games.

Oh you can figure it out, can you? How? Based on your sole Crysis benchmark and some 3DMark synthetics that mean nothing in the real world?

Others can analyze things too, but also know the value of looking at range of real-world results to come up with a truer picture. That means looking outside of 3DMark and a single Crysis benchmark before trying to make predictions about trends in the real world.

Techreport tested and gave an analysis which I come to agree with when they did the pixel fill test because I've done these kind of testings as well.

Again, those 3DMark results generally have absolutely no correlation to reality. A synthetic fillrate test tells you absolutely nothing about how the card behaves when other parts of its pipeline are being strained, as is the norm when running actual games.

Like I said. You finally agreed with what I've been saying all this time and that you were wrong after 2 years of BS you've been giving me the run around.

No, I agreed that in the specific case of the 8800 Ultra I was surprised that the core showed a bigger benefit. I also pointed out that the shader was second (without AA), and that changes to memory clocks still had an effect (with or without AA) despite you labeling the card as bandwidth happy akin to the 2900XT.

I also then repeatedly pointed out the example of the X1900 vs the X1800 where the former was basically the same as the latter, just with three times the shading capability. I then repeatedly linked to benchmarks of that era showing large performance gains in a range of titles.

I'll let your tests and my tests do the talking when you realize it for yourself in another 5 years!!!

Yep, much like the 260+ tests I linked to which showed the shader being slightly ahead of core and memory without AA, and showed the part being quite balanced overall despite your reference to it being bandwidth happy.

And yes, I do feel like running some new benchmarks soon.

 

[AzN]

Originally posted by: BFG10K
No, I agreed that in the specific case of the 8800 Ultra I was surprised that the core showed a bigger benefit. I also pointed out that the shader was second (without AA), and that changes to memory clocks still had an effect (with or without AA) despite you labeling the card as bandwidth happy akin to the 2900XT.

I'm just going to say this last thing and I'm out of this thread.

Considering you said Shader makes the most difference in all cards including g92, ultra, or whatever in late 2007 only to discover it didn't make the most difference in early 2008 when you tested your ultra. So you agree with me now in 2009. ROFL...

How did I know fillrate and right combination of bandwidth makes the most difference because I correctly tested them while you cited internet websites to agree with you which you follow mindlessly.

I'm not even going to debate with you any further because all you ever do is give out rhetorical responses peddling away from the subject at hand.

Some people really need to practice what they preach.