ATI 4xxx Series Thread

[SlowSpyder]

Originally posted by: Crisium
Has it been known yet if the 4800 series will be able to crossfire with the 3800 series? If yes, will it work in This motherboard?

Even if you could I'm not sure that you'd want to on that board. If I'm reading the specs right that board has a 16x PCIE slot as well as a 4x PCIE slot. I would think that 4x would be a bottle neck on either a 3870 or a 48x0 card.

 

[SickBeast]

Originally posted by: Extelleron
You make it sound as if RV770 will not have enough texture power.... I don't think that is true. 32 TMU at 750MHz is a significant amount of texture power. G80 did fine with 32 TMU @ 575MHz, didn't it? So why is 32 TMU at a much higher frequency now not enough? Pixel wise RV670 had enough power. HD 3870 has much more pixel power than even the 9800GTX and rumors seem to suggest that the RBE's in RV770 will be improved a bit from RV670.

32TMU at 750mhz is approximately the same amount of texture power that a mildly overclocked 8800GT has. Speaking of G80, it came out a year and a half ago. I would expect a true 'next gen' GPU to double it in all aspects at this point.

The one positive thing that I can see in the 4850/70 specs is that there is a very good chance that the cards will perform twice as fast as a 3850/70. By doubling the TMUs, hopefully that will alleviate one of the main bottlenecks that most people have pointed out in the R600 and later architectures from AMD. The other is the whole 'AA done by the shaders' debacle, but rumor has it that has been fixed as well.

As for the rest of your post, it's my opinion that the sheer number of shaders is the problem with R600, not the total amount of shading power. It seems that higher clocked, more powerful shaders are far more effective than a whole pile of weak ones.

 

[Extelleron]

Originally posted by: SickBeast

Originally posted by: Extelleron
You make it sound as if RV770 will not have enough texture power.... I don't think that is true. 32 TMU at 750MHz is a significant amount of texture power. G80 did fine with 32 TMU @ 575MHz, didn't it? So why is 32 TMU at a much higher frequency now not enough? Pixel wise RV670 had enough power. HD 3870 has much more pixel power than even the 9800GTX and rumors seem to suggest that the RBE's in RV770 will be improved a bit from RV670.

32TMU at 750mhz is approximately the same amount of texture power that a mildly overclocked 8800GT has. Speaking of G80, it came out a year and a half ago. I would expect a true 'next gen' GPU to double it in all aspects at this point.

The one positive thing that I can see in the 4850/70 specs is that there is a very good chance that the cards will perform twice as fast as a 3850/70. By doubling the TMUs, hopefully that will alleviate one of the main bottlenecks that most people have pointed out in the R600 and later architectures from AMD. The other is the whole 'AA done by the shaders' debacle, but rumor has it that has been fixed as well.

As for the rest of your post, it's my opinion that the sheer number of shaders is the problem with R600, not the total amount of shading power. It seems that higher clocked, more powerful shaders are far more effective than a whole pile of weak ones.

That might be true, but the 8800 Ultra is the most powerful single-GPU card you can buy at this point, and it gets along quite fine with 32 TMUs @ 612MHz. It doesn't have any problem beating a 9800GTX that has "64" TMUs clocked at 675MHz (Not a completely fair comparison, the difference is G80 had a 1:2 texture address: filtering ratio, G92 has 1:1 like G84). I do not believe that 32 TMUs @ 750MHz is going to be a serious bottleneck for RV770. Remember RV770 is not the "next gen" card.... R700 is. RV770 is a midrange card designed for the <$300 market. I don't think anyone can complain about it having 22% more texture power than the 8800 Ultra.

The problem with R600 is not the number of shading processors nor is it the clock. R600's 5-way shading processors have 4 "weak" SPs and a single "fat" SP capable of more complex calculations. This setup is not ideal for all games, and this causes some of the resources to be underutilized. If you get an application that is coded to work well with the R600 architecture, I am sure you can see great things with it. But this is not always the case, and that is why you see R600 underperforming often.

This is not to say that the R600 architecture is bad or horribly inefficient. If RV770 is really 800SP + 32 TMU + 16 buffed-up RBE's and yet only 34% larger than RV670, then that shows just how efficient the R600 architecture is in terms of utilizing minimal die area. Look at G92 -> GT200, you go from 324mm^2 to 576mm^2, yet all you get is less than double the SP's, only 16 more TMUs, and more ROPs (whether or not G92 actually has 16 or 24 ROPs we will never know exactly). And that 576mm^2 doesn't include the display logic, which is a seperate chip with GT200 parts like it was with G80. So comparing RV670 -> RV770 to G92 -> GT200, AMD increases die size by 34%, nVidia increases die size by 78%+, yet both will supposedly see ~2x+ performance from the previous architecture.

 

[allies]

Originally posted by: Extelleron

Originally posted by: allies

Originally posted by: Extelleron

Originally posted by: allies
I don't think there's anyway in hell that the R770 has 800 SPs. That "AMD" slide looks photoshopped.

What makes you think it will not have 800SPs?

The die size is enough to accomodate it. AMD's SPs do not take up much die area.

Because we've heard from EVERYWHERE that it's 480, including AMD. Yet this slide leaks out that is obviously photoshopped claiming 800.

When did AMD ever tell us anything about RV770, except that it would use GDDR5 memory?

The guy who runs Nordic Hardware said that the person that slide came from is highly reliable and has been 100% accurate in the past. He says that all of his contacts in the East have said 800SP's from the beginning, and the 480SP rumor originated from a German website.

And the slide is not obviously photoshopped. It is possible that it is, but it is not exactly obvious. Anyway, doesn't the previously rumored (and from CCC shots, apparently accurate) 625MHz core clock for the HD 4850 equalling exactly 1 TFLOP seem too perfect for it not to be true?

Link for the bolded please.

 

[taltamir]

Originally posted by: Extelleron

Originally posted by: SickBeast

Originally posted by: Extelleron

Originally posted by: allies

Originally posted by: Extelleron

Originally posted by: allies
I don't think there's anyway in hell that the R770 has 800 SPs. That "AMD" slide looks photoshopped.

What makes you think it will not have 800SPs?

The die size is enough to accomodate it. AMD's SPs do not take up much die area.

Because we've heard from EVERYWHERE that it's 480, including AMD. Yet this slide leaks out that is obviously photoshopped claiming 800.

When did AMD ever tell us anything about RV770, except that it would use GDDR5 memory?

The guy who runs Nordic Hardware said that the person that slide came from is highly reliable and has been 100% accurate in the past. He says that all of his contacts in the East have said 800SP's from the beginning, and the 480SP rumor originated from a German website.

And the slide is not obviously photoshopped. It is possible that it is, but it is not exactly obvious. Anyway, doesn't the previously rumored (and from CCC shots, apparently accurate) 625MHz core clock for the HD 4850 equalling exactly 1 TFLOP seem too perfect for it not to be true?

R600 also had a ridiculous number of SPs for its time, and we all know how that turned out. :roll:

Beyond a certain point, SPs become useless. Several of us verified this a few months ago by disabling SPs on our graphics cards. Up to a certain point, there was absolutely zero performance impact in doing so.

If AMD did in fact dedicate enough die space for 800SPs, to me that shows that they lack proper direction in terms of the type of GPU power that current games require. With that many SPs and so few TMUs and ROPs, the leaked specs make the card look terribly unbalanced IMO.

R600 did not have a rediculous number of stream processors. It had 64 5-way processors that marketing makes sound crazy, with the 320SP figure. Even if you say that R600 is 320SP, they are clocked at only 745MHz compared to 1.35GHz for G80. R600 did not have a rediculous amount of shading power compared to G80. Vertex shading, maybe. But not pixel.

You make it sound as if RV770 will not have enough texture power.... I don't think that is true. 32 TMU at 750MHz is a significant amount of texture power. G80 did fine with 32 TMU @ 575MHz, didn't it? So why is 32 TMU at a much higher frequency now not enough? Pixel wise RV670 had enough power. HD 3870 has much more pixel power than even the 9800GTX and rumors seem to suggest that the RBE's in RV770 will be improved a bit from RV670.

The point is, SP's do not take up much die space at all. So there is really no reason not to have 800SP. Having a significant amount of shading power will improve performance in games that benefit from it and also will make the card great for GPGPU uses such as F@H, etc; TMUs don't do much for non-gaming purposes. Obviously gaming is 1st priority, but I don't think anyone will be complaining about gaming performance on a HD 4870. Except for Crysis @ VH 1920x1200++, a $300 HD 4870 will be the best buy on the market (from what I am speculating and from what I am hearing). If you have a CF mobo, HD 4850 CF for $400 will be a great deal too.

excellent points. Funny that you mention crysis, as it is the only game where shading was ALWAYS the limiting factor for power. The tests showing shader power having no impact were with other games, crysis eats shaders up.

 

[Extelleron]

Originally posted by: allies

Originally posted by: Extelleron

Originally posted by: allies

Originally posted by: Extelleron

Originally posted by: allies
I don't think there's anyway in hell that the R770 has 800 SPs. That "AMD" slide looks photoshopped.

What makes you think it will not have 800SPs?

The die size is enough to accomodate it. AMD's SPs do not take up much die area.

Because we've heard from EVERYWHERE that it's 480, including AMD. Yet this slide leaks out that is obviously photoshopped claiming 800.

When did AMD ever tell us anything about RV770, except that it would use GDDR5 memory?

The guy who runs Nordic Hardware said that the person that slide came from is highly reliable and has been 100% accurate in the past. He says that all of his contacts in the East have said 800SP's from the beginning, and the 480SP rumor originated from a German website.

And the slide is not obviously photoshopped. It is possible that it is, but it is not exactly obvious. Anyway, doesn't the previously rumored (and from CCC shots, apparently accurate) 625MHz core clock for the HD 4850 equalling exactly 1 TFLOP seem too perfect for it not to be true?

Link for the bolded please.

http://www.xtremesystems.org/f...3043476&postcount=1590

Obviously he could be wrong, but I think NH would take a beating for saying he was so certain two weeks before launch and then being dead wrong.

 

[ronnn]

Originally posted by: SickBeast

R600 also had a ridiculous number of SPs for its time, and we all know how that turned out. :roll:

Beyond a certain point, SPs become useless. Several of us verified this a few months ago by disabling SPs on our graphics cards. Up to a certain point, there was absolutely zero performance impact in doing so.

If AMD did in fact dedicate enough die space for 800SPs, to me that shows that they lack proper direction in terms of the type of GPU power that current games require. With that many SPs and so few TMUs and ROPs, the leaked specs make the card look terribly unbalanced IMO.

Now this is quite different than saying this is a photoshop and a fake leak. How the card performs is I agree, going to be the issue. All we seem to have so far is a bit of 3d mock - ATi has been very secretive this time around. Seems to me that there are more hints of nice card from fairly reliable sources. Hard to decipher with all the vrzone noise. Wish they would not speculate on amd - as they seem to only have connections in the other camp.

 

[TC91]

Imagine it was 800sp's with a separate shader clock, that would pretty cool to play with.

 

[SickBeast]

Originally posted by: ronnn

Originally posted by: SickBeast

R600 also had a ridiculous number of SPs for its time, and we all know how that turned out. :roll:

Beyond a certain point, SPs become useless. Several of us verified this a few months ago by disabling SPs on our graphics cards. Up to a certain point, there was absolutely zero performance impact in doing so.

If AMD did in fact dedicate enough die space for 800SPs, to me that shows that they lack proper direction in terms of the type of GPU power that current games require. With that many SPs and so few TMUs and ROPs, the leaked specs make the card look terribly unbalanced IMO.

Now this is quite different than saying this is a photoshop and a fake leak. How the card performs is I agree, going to be the issue. All we seem to have so far is a bit of 3d mock - ATi has been very secretive this time around. Seems to me that there are more hints of nice card from fairly reliable sources. Hard to decipher with all the vrzone noise. Wish they would not speculate on amd - as they seem to only have connections in the other camp.

The thing is, the 3DMock performance actually puts the card pretty close to where it should be based on its specs. One could hope for 20-30% more than the result shows, but I'm fairly confident that the 4850 will be in the same ballpark as an 8800GT.

There has been a fair bit of positive buzz surrounding the 4850/70. The 3870 also received positive reviews, but really that card never interested me.

As far as I'm concerned, AMD needs to build a new GPU from the ground up. By all acounts, their new GPUs are simply re-incarnations of their older stuff.

 

[Extelleron]

Originally posted by: SickBeast

Originally posted by: ronnn

Originally posted by: SickBeast

R600 also had a ridiculous number of SPs for its time, and we all know how that turned out. :roll:

Beyond a certain point, SPs become useless. Several of us verified this a few months ago by disabling SPs on our graphics cards. Up to a certain point, there was absolutely zero performance impact in doing so.

If AMD did in fact dedicate enough die space for 800SPs, to me that shows that they lack proper direction in terms of the type of GPU power that current games require. With that many SPs and so few TMUs and ROPs, the leaked specs make the card look terribly unbalanced IMO.

Now this is quite different than saying this is a photoshop and a fake leak. How the card performs is I agree, going to be the issue. All we seem to have so far is a bit of 3d mock - ATi has been very secretive this time around. Seems to me that there are more hints of nice card from fairly reliable sources. Hard to decipher with all the vrzone noise. Wish they would not speculate on amd - as they seem to only have connections in the other camp.

The thing is, the 3DMock performance actually puts the card pretty close to where it should be based on its specs. One could hope for 20-30% more than the result shows, but I'm fairly confident that the 4850 will be in the same ballpark as an 8800GT.

There has been a fair bit of positive buzz surrounding the 4850/70. The 3870 also received positive reviews, but really that card never interested me.

As far as I'm concerned, AMD needs to build a new GPU from the ground up. By all acounts, their new GPUs are simply re-incarnations of their older stuff.

How are you thinking that it will be close to an 8800GT?

Compared to HD 3870, HD 4850 should offer 2.02x more shading power & 1.61x more texture power.

The only thing that I can see being a clear bottleneck is memory bandwidth. It's possible that the 2000MHz GDDR3 the 4850 will use will limit its performance. But even so, I think we can confidently say the HD 4850 will be 50% faster in every situation than the HD 3870, if not much more. There are very few, if any, situations where the 8800GT is even close to 50% faster than the HD 3870. On average it is maybe 15-20% faster.

You are talking about AMD using the same architecture, what is nVidia doing? GT200 is more of the same.

If you consider improvement from last year to this year, AMD is way in the lead. Compared to the performance of R600 just a year ago, AMD will have a part 3-4x faster in R700. nVidia, meanwhile, is going to struggle to get 2x 8800 Ultra performance well over a year after G80 came out.

AMD does not need a new architecture. R600 is a great architecture; it is highly scalable, does not take up much die area, and performs great in a number of games (especially UE3 games, which dominate the market). And they will have the performance lead or at least parity with R700. I thought AMD was down for the count after R600, but clearly 2007 is a different year for AMD compared to 2006.

I do not know what nVidia has in mind after GT200, but if it is more of single-GPU, then I would be worried if I were them. GT200b 55nm will likely be out late this year or early next year, but by that time AMD could be refreshing RV770 on either 45nm or 40nm. That is the great thing about staying with a reasonably sized GPU; you can move to the latest manufacturing processes even when they are risky, something nVidia cannot do with a huge GPU like GT200.

 

[taltamir]

Originally posted by: SickBeast

Originally posted by: ronnn

Originally posted by: SickBeast

R600 also had a ridiculous number of SPs for its time, and we all know how that turned out. :roll:

Beyond a certain point, SPs become useless. Several of us verified this a few months ago by disabling SPs on our graphics cards. Up to a certain point, there was absolutely zero performance impact in doing so.

If AMD did in fact dedicate enough die space for 800SPs, to me that shows that they lack proper direction in terms of the type of GPU power that current games require. With that many SPs and so few TMUs and ROPs, the leaked specs make the card look terribly unbalanced IMO.

Now this is quite different than saying this is a photoshop and a fake leak. How the card performs is I agree, going to be the issue. All we seem to have so far is a bit of 3d mock - ATi has been very secretive this time around. Seems to me that there are more hints of nice card from fairly reliable sources. Hard to decipher with all the vrzone noise. Wish they would not speculate on amd - as they seem to only have connections in the other camp.

The thing is, the 3DMock performance actually puts the card pretty close to where it should be based on its specs. One could hope for 20-30% more than the result shows, but I'm fairly confident that the 4850 will be in the same ballpark as an 8800GT.

There has been a fair bit of positive buzz surrounding the 4850/70. The 3870 also received positive reviews, but really that card never interested me.

As far as I'm concerned, AMD needs to build a new GPU from the ground up. By all acounts, their new GPUs are simply re-incarnations of their older stuff.

everyone is using older stuff in their architecture... rebuilding from scratch is an insane endeavor that will lead to nowhere fast.

 

[SickBeast]

Originally posted by: taltamir
everyone is using older stuff in their architecture... rebuilding from scratch is an insane endeavor that will lead to nowhere fast.

With the R600-based stuff as borked as it is, they should be more drastically revamping the architecture if not starting over. The shader architecture is horribly inefficient, the GPU lacks TMUs, it cannot perform proper AA, plus they wasted a ton of die space making the thing 512-bit unnecessarily.

NV started from scratch after their 5900-series cards and came out with the 6800-series which was much better. The G80 is also a new architecture, and again it performs quite well.

From what I understand, any semiconductor company pretty much has to 'start from scratch' every time there is a major die shrink involved. Usually they make a part, then perform a minor die shrink refresh, then make a new part altogether.

By saying they should start over, I obviously don't mean that they should throw away everything they know about GPU design. I simply mean that the parts need to be completely revamped.

 

[SickBeast]

Originally posted by: Extelleron

Originally posted by: SickBeast

Originally posted by: ronnn

Originally posted by: SickBeast

R600 also had a ridiculous number of SPs for its time, and we all know how that turned out. :roll:

Beyond a certain point, SPs become useless. Several of us verified this a few months ago by disabling SPs on our graphics cards. Up to a certain point, there was absolutely zero performance impact in doing so.

If AMD did in fact dedicate enough die space for 800SPs, to me that shows that they lack proper direction in terms of the type of GPU power that current games require. With that many SPs and so few TMUs and ROPs, the leaked specs make the card look terribly unbalanced IMO.

Now this is quite different than saying this is a photoshop and a fake leak. How the card performs is I agree, going to be the issue. All we seem to have so far is a bit of 3d mock - ATi has been very secretive this time around. Seems to me that there are more hints of nice card from fairly reliable sources. Hard to decipher with all the vrzone noise. Wish they would not speculate on amd - as they seem to only have connections in the other camp.

The thing is, the 3DMock performance actually puts the card pretty close to where it should be based on its specs. One could hope for 20-30% more than the result shows, but I'm fairly confident that the 4850 will be in the same ballpark as an 8800GT.

There has been a fair bit of positive buzz surrounding the 4850/70. The 3870 also received positive reviews, but really that card never interested me.

As far as I'm concerned, AMD needs to build a new GPU from the ground up. By all acounts, their new GPUs are simply re-incarnations of their older stuff.

How are you thinking that it will be close to an 8800GT?

Compared to HD 3870, HD 4850 should offer 2.02x more shading power & 1.61x more texture power.

The only thing that I can see being a clear bottleneck is memory bandwidth. It's possible that the 2000MHz GDDR3 the 4850 will use will limit its performance. But even so, I think we can confidently say the HD 4850 will be 50% faster in every situation than the HD 3870, if not much more. There are very few, if any, situations where the 8800GT is even close to 50% faster than the HD 3870. On average it is maybe 15-20% faster.

You are talking about AMD using the same architecture, what is nVidia doing? GT200 is more of the same.

If you consider improvement from last year to this year, AMD is way in the lead. Compared to the performance of R600 just a year ago, AMD will have a part 3-4x faster in R700. nVidia, meanwhile, is going to struggle to get 2x 8800 Ultra performance well over a year after G80 came out.

AMD does not need a new architecture. R600 is a great architecture; it is highly scalable, does not take up much die area, and performs great in a number of games (especially UE3 games, which dominate the market). And they will have the performance lead or at least parity with R700. I thought AMD was down for the count after R600, but clearly 2007 is a different year for AMD compared to 2006.

I do not know what nVidia has in mind after GT200, but if it is more of single-GPU, then I would be worried if I were them. GT200b 55nm will likely be out late this year or early next year, but by that time AMD could be refreshing RV770 on either 45nm or 40nm. That is the great thing about staying with a reasonably sized GPU; you can move to the latest manufacturing processes even when they are risky, something nVidia cannot do with a huge GPU like GT200.

On paper, it's very comparable to an 8800GT within 30% or so.

- TMUs are the same
- clockspeed is the same (in the case of the 4850)
- shading power is a mystery; hopefully AMD has used a more efficient design here (doubtful - it looks like they want 800 crappy shaders on the spec sheet to fool uninformed consumers)
- 16 ROPs vs. 24 or so on the 8800GT
- similar cooling designs

The 4870 may in fact be 30% faster if they can clock it to 800mhz. While this is significant, to me it's a hollow victory in light of the fact that G80/G92 have been on the market for quite some time.

As for R600, see my post above.

As for GT200 being a 'refresh' of the G80/G92, I'm pretty sure it's not.

I just find it frustrating the AMD/ATI have allowed the AA bug on R600 to carry forward through several product cycles. There are a whole pile of games that cannot even run AA on AMD hardware.

Hopefully the new AMD parts will perform twice as fast as a 3870 like they say they will. To me if they can do that with only 34% more transistors it further proves how messed up R600 was. Their entire product line, CPUs included, has seemed rushed and incomplete ever since the merger. :thumbsdown:

 

[allies]

Originally posted by: SickBeast

Originally posted by: taltamir
everyone is using older stuff in their architecture... rebuilding from scratch is an insane endeavor that will lead to nowhere fast.

With the R600-based stuff as borked as it is, they should be more drastically revamping the architecture if not starting over. The shader architecture is horribly inefficient, the GPU lacks TMUs, it cannot perform proper AA, plus they wasted a ton of die space making the thing 512-bit unnecessarily.

NV started from scratch after their 5900-series cards and came out with the 6800-series which was much better. The G80 is also a new architecture, and again it performs quite well.

From what I understand, any semiconductor company pretty much has to 'start from scratch' every time there is a major die shrink involved. Usually they make a part, then to a minor die shrink refresh, then make a new part altogether.

By saying they should start over, I obviously don't mean that they should throw away everything they know about GPU design. I simply mean that the parts need to be completely revamped.

The R770 already addresses most of the points you make:

• (rumor)Revamped AA
  2x TMUs
  More shaders (maybe tweaking involve)
  256mbit bus with GDDR5

What more do you want again

 

[Janooo]

Originally posted by: SickBeast

Originally posted by: taltamir
everyone is using older stuff in their architecture... rebuilding from scratch is an insane endeavor that will lead to nowhere fast.

With the R600-based stuff as borked as it is, they should be more drastically revamping the architecture if not starting over. The shader architecture is horribly inefficient, the GPU lacks TMUs, it cannot perform proper AA, plus they wasted a ton of die space making the thing 512-bit unnecessarily.

NV started from scratch after their 5900-series cards and came out with the 6800-series which was much better. The G80 is also a new architecture, and again it performs quite well.

From what I understand, any semiconductor company pretty much has to 'start from scratch' every time there is a major die shrink involved. Usually they make a part, then perform a minor die shrink refresh, then make a new part altogether.

By saying they should start over, I obviously don't mean that they should throw away everything they know about GPU design. I simply mean that the parts need to be completely revamped.

If they took 770 and made it the size of the G200 what do you think the specs would looked like?
Performance? I think the G200 wouldn't have a chance if 700 with the multi GPU overhead is going to end up faster.

 

[SickBeast]

Originally posted by: allies

Originally posted by: SickBeast

Originally posted by: taltamir
everyone is using older stuff in their architecture... rebuilding from scratch is an insane endeavor that will lead to nowhere fast.

With the R600-based stuff as borked as it is, they should be more drastically revamping the architecture if not starting over. The shader architecture is horribly inefficient, the GPU lacks TMUs, it cannot perform proper AA, plus they wasted a ton of die space making the thing 512-bit unnecessarily.

NV started from scratch after their 5900-series cards and came out with the 6800-series which was much better. The G80 is also a new architecture, and again it performs quite well.

From what I understand, any semiconductor company pretty much has to 'start from scratch' every time there is a major die shrink involved. Usually they make a part, then to a minor die shrink refresh, then make a new part altogether.

By saying they should start over, I obviously don't mean that they should throw away everything they know about GPU design. I simply mean that the parts need to be completely revamped.

The R770 already addresses most of the points you make:

• (rumor)Revamped AA
  2x TMUs
  More shaders (maybe tweaking involve)
  256mbit bus with GDDR5

What more do you want again

I want proper shaders, not 800 crappy ones! :evil:

I suppose that's my main complaint with the recent AMD architectures. Their shader design is terrible. If they're simply adding 500 more crappy and inefficient shaders, I really don't see the point.

It just seems like it has taken AMD a long time to work the bugs out of R600. IMO a completely revamped GPU with better shaders and proper AA would have been the way to go. Hopefully the 4850/70 are just that. I'm not holding my breath. :beer:

 

[taltamir]

those shaders are each smaller then the more efficient nvidia shaders... maybe they should revamp the shaders, but that wouldn't be a new design, just new shaders. (improved ones actually... nvidia says it improved its shaders by 50%)

i understand what you are saying though, a major revamp might be in order... except they have different goals. They are majorly revamping it so that they could put GPUs and CPUs together.

Originally posted by: allies

• (rumor)Revamped AA
  2x TMUs
  More shaders (maybe tweaking involve)
  256mbit bus with GDDR5

What more do you want again

And the revamped CF communications...

 

[Janooo]

Originally posted by: SickBeast

Originally posted by: allies

Originally posted by: SickBeast

Originally posted by: taltamir
everyone is using older stuff in their architecture... rebuilding from scratch is an insane endeavor that will lead to nowhere fast.

With the R600-based stuff as borked as it is, they should be more drastically revamping the architecture if not starting over. The shader architecture is horribly inefficient, the GPU lacks TMUs, it cannot perform proper AA, plus they wasted a ton of die space making the thing 512-bit unnecessarily.

NV started from scratch after their 5900-series cards and came out with the 6800-series which was much better. The G80 is also a new architecture, and again it performs quite well.

From what I understand, any semiconductor company pretty much has to 'start from scratch' every time there is a major die shrink involved. Usually they make a part, then to a minor die shrink refresh, then make a new part altogether.

By saying they should start over, I obviously don't mean that they should throw away everything they know about GPU design. I simply mean that the parts need to be completely revamped.

The R770 already addresses most of the points you make:

• (rumor)Revamped AA
  2x TMUs
  More shaders (maybe tweaking involve)
  256mbit bus with GDDR5

What more do you want again

I want proper shaders, not 800 crappy ones! :evil:

I suppose that's my main complaint with the recent AMD architectures. Their shader design is terrible. If they're simply adding 500 more crappy and inefficient shaders, I really don't see the point.

It just seems like it has taken AMD a long time to work the bugs out of R600. IMO a completely revamped GPU with better shaders and proper AA would have been the way to go. Hopefully the 4850/70 are just that. I'm not holding my breath. :beer:

Please, check this tread and tell me if you still think they are crappy.

 

[SickBeast]

Originally posted by: Janooo

Originally posted by: SickBeast

Originally posted by: taltamir
everyone is using older stuff in their architecture... rebuilding from scratch is an insane endeavor that will lead to nowhere fast.

With the R600-based stuff as borked as it is, they should be more drastically revamping the architecture if not starting over. The shader architecture is horribly inefficient, the GPU lacks TMUs, it cannot perform proper AA, plus they wasted a ton of die space making the thing 512-bit unnecessarily.

NV started from scratch after their 5900-series cards and came out with the 6800-series which was much better. The G80 is also a new architecture, and again it performs quite well.

From what I understand, any semiconductor company pretty much has to 'start from scratch' every time there is a major die shrink involved. Usually they make a part, then perform a minor die shrink refresh, then make a new part altogether.

By saying they should start over, I obviously don't mean that they should throw away everything they know about GPU design. I simply mean that the parts need to be completely revamped.

If they took 770 and made it the size of the G200 what do you think the specs would looked like?
Performance? I think the G200 wouldn't have a chance if 700 with the multi GPU overhead is going to end up faster.

AMD could double their specs with that much die size (although 1600 shaders would be a little out there IMO).

Even if the R700 turns out faster, it uses two chips and effectively the G200 still wins if they are in SLI.

Using multiple GPUs causes problems. It wastes 1/2 of the texture memory that you pay for, causes stuttering, and often leaves you stuck running on a single GPU because the game you want to play doesn't have an SLI/CF profile.

 

[Janooo]

Originally posted by: SickBeast

Originally posted by: Janooo

Originally posted by: SickBeast

Originally posted by: taltamir
everyone is using older stuff in their architecture... rebuilding from scratch is an insane endeavor that will lead to nowhere fast.

With the R600-based stuff as borked as it is, they should be more drastically revamping the architecture if not starting over. The shader architecture is horribly inefficient, the GPU lacks TMUs, it cannot perform proper AA, plus they wasted a ton of die space making the thing 512-bit unnecessarily.

NV started from scratch after their 5900-series cards and came out with the 6800-series which was much better. The G80 is also a new architecture, and again it performs quite well.

From what I understand, any semiconductor company pretty much has to 'start from scratch' every time there is a major die shrink involved. Usually they make a part, then perform a minor die shrink refresh, then make a new part altogether.

By saying they should start over, I obviously don't mean that they should throw away everything they know about GPU design. I simply mean that the parts need to be completely revamped.

If they took 770 and made it the size of the G200 what do you think the specs would looked like?
Performance? I think the G200 wouldn't have a chance if 700 with the multi GPU overhead is going to end up faster.

AMD could double their specs with that much die size (although 1600 shaders would be a little out there IMO).

Even if the R700 turns out faster, it uses two chips and effectively the G200 still wins if they are in SLI.

Using multiple GPUs causes problems. It wastes 1/2 of the texture memory that you pay for, causes stuttering, and often leaves you stuck running on a single GPU because the game you want to play doesn't have an SLI/CF profile.

That was true for 3870X2. We will have to wait for the 4870X2 if it's going to be same.

 

[taltamir]

its a bit naive to assume AMD will be pulling a rabbit THAT fat out of the hat. They might, there are some rumors they might, but I always hear rumors that amount to nothing. the current tech, the one that is out and can be bought, has those problems and more.

 

[BFG10K]

AMD does not need a new architecture. R600 is a great architecture; it is highly scalable, does not take up much die area, and performs great in a number of games (especially UE3 games, which dominate the market).

It doesn't perform great except in rare games such as Call of Juarez, and only when AA is disabled.

The architecture currently has serious deficiencies especially in regard to shader performance where in the worst case scenario you only have one fifth of the quoted shader power available. And if that wasn?t bad enough AA taxes what little shader power is left.

nVidia?s shaders are far more generic and flexible plus they run at a massive clock speed advantage. That and they don?t deal with AA.

Hopefully ATi will address most of the issues and hopefully 800 shaders will be enough to offset the remaining issues.

 

[Hauk]

I found a pretty nice pic of the 4870. If you look carefully, you'll see the heatsink design through the somewhat transparent housing.

It appears to have a copper block over the gpu from which two large diameter heatpipes connect to the latter half of long aluminium fins.

 

[BFG10K]

Using multiple GPUs causes problems. It wastes 1/2 of the texture memory that you pay for,

I'm not sure what you mean here. Do you mean it wastes the VRAM in general, as in you pay for 2 x 512 MB cards but you only effectively get 512 MB?

I just find it frustrating the AMD/ATI have allowed the AA bug on R600 to carry forward through several product cycles. There are a whole pile of games that cannot even run AA on AMD hardware.

What AA bug? The only AA ?bug? is the fact that shader resolve is done through shaders and not through ROPs which makes it slow.

The compatibility of AA in games is a driver issue as said games often use non-standard rendering (e.g. FP HDR, deferred rendering, etc) and hence need driver workarounds. nVidia puts more effort into their drivers so they get AA working faster and into more games.

This is quite disappointing from ATi?s standpoint as their AA is basically a post-filter which should make it much easier to enable than traditional hardware methods working earlier in the pipeline.

 

[Extelleron]

Originally posted by: SickBeast

Originally posted by: Extelleron

Originally posted by: SickBeast

Originally posted by: ronnn

Originally posted by: SickBeast

R600 also had a ridiculous number of SPs for its time, and we all know how that turned out. :roll:

Beyond a certain point, SPs become useless. Several of us verified this a few months ago by disabling SPs on our graphics cards. Up to a certain point, there was absolutely zero performance impact in doing so.

If AMD did in fact dedicate enough die space for 800SPs, to me that shows that they lack proper direction in terms of the type of GPU power that current games require. With that many SPs and so few TMUs and ROPs, the leaked specs make the card look terribly unbalanced IMO.

Now this is quite different than saying this is a photoshop and a fake leak. How the card performs is I agree, going to be the issue. All we seem to have so far is a bit of 3d mock - ATi has been very secretive this time around. Seems to me that there are more hints of nice card from fairly reliable sources. Hard to decipher with all the vrzone noise. Wish they would not speculate on amd - as they seem to only have connections in the other camp.

The thing is, the 3DMock performance actually puts the card pretty close to where it should be based on its specs. One could hope for 20-30% more than the result shows, but I'm fairly confident that the 4850 will be in the same ballpark as an 8800GT.

There has been a fair bit of positive buzz surrounding the 4850/70. The 3870 also received positive reviews, but really that card never interested me.

As far as I'm concerned, AMD needs to build a new GPU from the ground up. By all acounts, their new GPUs are simply re-incarnations of their older stuff.

How are you thinking that it will be close to an 8800GT?

Compared to HD 3870, HD 4850 should offer 2.02x more shading power & 1.61x more texture power.

The only thing that I can see being a clear bottleneck is memory bandwidth. It's possible that the 2000MHz GDDR3 the 4850 will use will limit its performance. But even so, I think we can confidently say the HD 4850 will be 50% faster in every situation than the HD 3870, if not much more. There are very few, if any, situations where the 8800GT is even close to 50% faster than the HD 3870. On average it is maybe 15-20% faster.

You are talking about AMD using the same architecture, what is nVidia doing? GT200 is more of the same.

If you consider improvement from last year to this year, AMD is way in the lead. Compared to the performance of R600 just a year ago, AMD will have a part 3-4x faster in R700. nVidia, meanwhile, is going to struggle to get 2x 8800 Ultra performance well over a year after G80 came out.

AMD does not need a new architecture. R600 is a great architecture; it is highly scalable, does not take up much die area, and performs great in a number of games (especially UE3 games, which dominate the market). And they will have the performance lead or at least parity with R700. I thought AMD was down for the count after R600, but clearly 2007 is a different year for AMD compared to 2006.

I do not know what nVidia has in mind after GT200, but if it is more of single-GPU, then I would be worried if I were them. GT200b 55nm will likely be out late this year or early next year, but by that time AMD could be refreshing RV770 on either 45nm or 40nm. That is the great thing about staying with a reasonably sized GPU; you can move to the latest manufacturing processes even when they are risky, something nVidia cannot do with a huge GPU like GT200.

On paper, it's very comparable to an 8800GT within 30% or so.

- TMUs are the same
- clockspeed is the same (in the case of the 4850)
- shading power is a mystery; hopefully AMD has used a more efficient design here (doubtful - it looks like they want 800 crappy shaders on the spec sheet to fool uninformed consumers)
- 16 ROPs vs. 24 or so on the 8800GT
- similar cooling designs

The 4870 may in fact be 30% faster if they can clock it to 800mhz. While this is significant, to me it's a hollow victory in light of the fact that G80/G92 have been on the market for quite some time.

As for R600, see my post above.

As for GT200 being a 'refresh' of the G80/G92, I'm pretty sure it's not.

I just find it frustrating the AMD/ATI have allowed the AA bug on R600 to carry forward through several product cycles. There are a whole pile of games that cannot even run AA on AMD hardware.

Hopefully the new AMD parts will perform twice as fast as a 3870 like they say they will. To me if they can do that with only 34% more transistors it further proves how messed up R600 was. Their entire product line, CPUs included, has seemed rushed and incomplete ever since the merger. :thumbsdown:

Well we're going to find out very soon who is right: http://www.xtremesystems.org/f...3049772&postcount=1773

BTW the 8800GT has 16 ROPs, same as HD 4850. G92 may or may not have 24 ROPs all together, but if so no released card has utilized them. The reason for that is just that they need 16 ROPs for a 256-bit bus; 24 ROPs would give a 384-bit bus and clearly they didn't want such an expensive bus with a "cheap" product like the 8800GT/S and the 9800GTX.

You are comparing specs directly as if G92 and RV770 were the exact same thing. You simply can't do that. As I said, HD 3870 is 64SP / 16 TMU and it is only ~20% slower than a 8800GT on average, despite the 8800GT being 112 SP / 56 TMU. The HD 4850 is 160 SP/ 32 TMU and slightly lower clock. Comparing it to the HD 3870, along with some rumored architectural improvements, I am pretty certain it will be 50% faster or more. And that would put it ahead of the 8800GT.

As for AA, it is actually not the problem on R600 at all. In fact in a number of situations R600 takes less / equal hit from AA compared to G80/G92 parts. The problem is AF. R600 parts take a huge hit from anisotropic filtering. Since AA/AF are almost always used together, many people assumed that AA was the problem.

When I say R600, I am talking about the architecture, not the chip. R600 as a chip was not very good; it was power hungry, huge, and didn't perform that well. But the architecture is very good IMO. The shaders are not very strong, but they are very small and you can fit a lot of them in the same space as nVidia's G80 SPs. There is nothing wrong with having 800SP's that are weak if they can compete with a lower number of nVidia shaders that fit in the same die space or less. Remember that G92b is rumored to be larger than RV770 on the same 55nm process; so that means (roughly) that 800 R600 SPs are smaller than 128 G80 SPs. Which one will perform better? I guarantee RV770 will have more shading power than G92.