R600 nice read

[Cooler]

The 1 gig Vram is not new. Some of the high end FireGL cards have had it last year. Its only natural that ATI would carry this over with GDDR4 on R600. Also there are rumors that the new internal memory control is 1024bits and 512bit external. (R520 was 512bit internal 256 external). The memory could be used for higher quality AF/AA. ATI always liked behead in this area . I dont think they would let G80 beat them at their own game.

 

[Howard]

Originally posted by: 5150Joker
The R600 is going to pwn the ****** out of nVidia. Hurry up AMD get this beast out there so I can get it plus a quad core system along with Crysis and other DX10 games!

What the hell?

 

[Cookie Monster]

You guys have to think that
a) using 1gb of GDDR4 memory is very expensive, especially the higher speed GDDR4 modules
b) 512bit makes PCB layout very complex i.e increasing the cost of the overall cost of the PCB
c) even though using GDDR4 can save power, 1gb seriously draws alot of power
d) pimping up the R600 with the latest and fastest memory could always be a problem in terms of supply

and all this affects the final price of the product.

 

[josh6079]

Originally posted by: Cookie Monster
You guys have to think that
a) using 1gb of GDDR4 memory is very expensive, especially the higher speed GDDR4 modules
b) 512bit makes PCB layout very complex i.e increasing the cost of the overall cost of the PCB
c) even though using GDDR4 can save power, 1gb seriously draws alot of power
d) pimping up the R600 with the latest and fastest memory could always be a problem in terms of supply

and all this affects the final price of the product.

So if it comes out around the same price as the G80 you'll know that either ATi is underpricing or Nvidia is overpricing, right? I mean, if the R600 sports more expensive gear than the G80 and costs relatively the same...

I see your point though. If they want $700 or $800 for it, no thanks, I'll get a G80 instead. My point is no matter what it costs them, they're going to have to price it relative to the competing GPU's. Otherwise, A LOT of itching entusiasts will turn it down.

 

[Wreckage]

Originally posted by: josh6079

Originally posted by: Cookie Monster
You guys have to think that
a) using 1gb of GDDR4 memory is very expensive, especially the higher speed GDDR4 modules
b) 512bit makes PCB layout very complex i.e increasing the cost of the overall cost of the PCB
c) even though using GDDR4 can save power, 1gb seriously draws alot of power
d) pimping up the R600 with the latest and fastest memory could always be a problem in terms of supply

and all this affects the final price of the product.

So if it comes out around the same price as the G80 you'll know that either ATi is underpricing or Nvidia is overpricing, right? I mean, if the R600 sports more expensive gear than the G80 and costs relatively the same...

That would also mean that NVIDIA could just offer a substantial price drop upon the R600 launch (here's to hoping anyway). Either way this delay can't be helping AMD any.

 

[TroubleM]

Pictures of the R600 core:
http://img218.imageshack.us/my.php?image=r600kn2.jpg
http://img220.imageshack.us/my.php?image=r600mkl9.jpg
http://img215.imageshack.us/my.php?image=r600nyc6.jpg
Back
http://www.iamxtreme.net/andre/R600.jpg
http://www.iamxtreme.net/andre/R600_3.jpg
http://www.iamxtreme.net/andre/R600_2.jpg

from http://www.beyond3d.com/forum/showthread.php?t=34676&page=27

 

[josh6079]

Either way this delay can't be helping AMD any.

Is it a delay? I never saw a press release from ATi saying that the R600 was going to be released in Q4 of 06.

 

[Wreckage]

Originally posted by: josh6079

Either way this delay can't be helping AMD any.

Is it a delay? I never saw a press release from ATi saying that the R600 was going to be released in Q4 of 06.

A matter of semantics I suppose. How about this.... Either way AMD releasing their DX10 part several months after NVIDIA can't be helping them any.

My use of the term "delay" is brought about by reading too many rumor sites I suppose. If you do a google search for R600 and delay you will find it is referenced quite often.

 

[jim1976]

Originally posted by: josh6079
Who's to say it won't need it, the GPU isn't even out. We have absolutely no idea kind of AA the R600 will push and nor what other features it may bring. For instance, what if it supported 16xSAA? That would need a huge amount of bandwitdth when using it with resolutions of 1600x1200 or higher.

I'm not saying it will use that kind of AA, but I just think it's too early to judge its needs for high bandwidth before we know what features it supports.

If you are gonna use 16xSS your fillrate will end up much earlier than you expect it to be. Now if you want to play a game that relies heavily on branching or texturing you can't possibly expect it to run at high resolutions in these settings..

Originally posted by: ronnn
I don't know enough about vista, unified architectures and upcoming games to dispute this. Still I have a feeling that r600 may rock in vista, while g80 wins in xp. Should lead to some good times with familiar arguments. With current games, not much reason for me to buy any $600 card - so I expect crysis and so forth to want more. :thumbsup:

I expect R600 to be a winner in almost all cases, but that's just me.. I'm just trying to make the point that ATI/AMD already have enough excess bandwith from GDDR4 even with this memory controller or a 384bit one. I just can't see why they would like to max out the memory bandwith when this could prove more helpful in a later refresh. I myself, like you though wish to see a 512bit bus in R600, only some fanatic kids wouldn't like it 😉 I just really wish that AMD/ATI have a really good reason to put the 512bit bus in there and that might have to do with something unfamiliar with what we already know so far from an architectural point of view.. Because as stated before, it will come as a costly investment and it will consume a lot more power(and this comes in contrast with the rumours that they want to make a smaller pcb that consumes less power and is more efficient). Whatever the case might be I sure hope they use 1GB. NOW THIS will be very helpful.

 

[jim1976]

Originally posted by: Cookie Monster
You guys have to think that
a) using 1gb of GDDR4 memory is very expensive, especially the higher speed GDDR4 modules
b) 512bit makes PCB layout very complex i.e increasing the cost of the overall cost of the PCB
c) even though using GDDR4 can save power, 1gb seriously draws alot of power
d) pimping up the R600 with the latest and fastest memory could always be a problem in terms of supply

and all this affects the final price of the product.

Good and valid points made.. We just have to wait and see what the truth is..

 

[josh6079]

If you are gonna use 16xSS your fillrate will end up much earlier than you expect it to be. Now if you want to play a game that relies heavily on branching or texturing you can't possibly expect it to run at high resolutions in these settings..

My previous example was just an analogy. All I'm saying is that until we know what algorithms the architecture uses it is pointless to debate about the necessity for such bandwidth.

A matter of semantics I suppose. How about this.... Either way AMD releasing their DX10 part several months after NVIDIA can't be helping them any.

It might if it is worth the wait. Imagine what would have happened if Nvidia wouldn't have caught their transistor problem earlier. Of course, I will agree that a lot is riding on this launch. The most recent ATi releases have been less than noteworthy.

My use of the term "delay" is brought about by reading too many rumor sites I suppose. If you do a google search for R600 and delay you will find it is referenced quite often.

It is referenced often because it finds all of the rumors you've mentioned. There hasn't been any official word yet that I know of.

We're just starting to see leaked pics and I assume benches are not far off either. I don't think ATi's R600 "delay" is hurting them anymore than AMD's "delay" is with its upcoming quad-cores.

Product cycles are just different sometimes, that's why I've had experience with several vendors; sometimes companies that normally do pretty good just don't have a product ready yet. When I got my X1900XTX, Nvidia didn't have anything that could answer to it. Same was the situation when I had my 7800GT's, nothing ATi had at the time could touch them.

 

[jim1976]

Originally posted by: josh6079
My previous example was just an analogy. All I'm saying is that until we know what algorithms the architecture uses it is pointless to debate about the necessity for such bandwidth.

I really hope I'm wrong and it turns out to be a 512bit controller with 1GB ram and on top of that it will be fully utilised. I just don't see it possible..

 

[Cooler]

Originally posted by: jim1976

Originally posted by: josh6079
My previous example was just an analogy. All I'm saying is that until we know what algorithms the architecture uses it is pointless to debate about the necessity for such bandwidth.

I really hope I'm wrong and it turns out to be a 512bit controller with 1GB ram and on top of that it will be fully utilised. I just don't see it possible..

It may allow for larger texture sizes greater than 4096x4096 that have never been allowed on home pcs before. With the more TMU in the R600 ATI is trying to fix its texturing weakness it had in the R520/R580.

 

[ronnn]

I find it interesting that amd seems to have leaked the pictures showing the pin count. Besides being good, inexpensive advertising - it does seem to confirm a 512 bit controller. Still this could turn out to be not so interesting - like the g80 mystery chip thing. Anyways I am confident that whatever it is - is not likely that ati would design a chip that was so unbalanced it could not compete with its current competition. As for cost of production, yields and profit margins - who knows. Is a tough environment right now to get people to spend big. Vista and a couple of games may change that though. As this generation is supposedly being designed for dx10.
This does suggest that it is getting closer to release - so maybe early January is accurate. :beer:

 

[Cookie Monster]

Originally posted by: Cooler

Originally posted by: jim1976

Originally posted by: josh6079
My previous example was just an analogy. All I'm saying is that until we know what algorithms the architecture uses it is pointless to debate about the necessity for such bandwidth.

I really hope I'm wrong and it turns out to be a 512bit controller with 1GB ram and on top of that it will be fully utilised. I just don't see it possible..

It may allow for larger texture sizes greater than 4096x4096 that have never been allowed on home pcs before. With the more TMU in the R600 ATI is trying to fix its texturing weakness it had in the R520/R580.

R580 and R520 never was weak in texturing. Efficency comes to mind, and at the very least these 16 decoupled TMUs were a match for G7x 24 TMUs. TMUs too have efficency and can vary in performance.

 

[josh6079]

Originally posted by: Cookie Monster

Originally posted by: Cooler

Originally posted by: jim1976

Originally posted by: josh6079
My previous example was just an analogy. All I'm saying is that until we know what algorithms the architecture uses it is pointless to debate about the necessity for such bandwidth.

I really hope I'm wrong and it turns out to be a 512bit controller with 1GB ram and on top of that it will be fully utilised. I just don't see it possible..

It may allow for larger texture sizes greater than 4096x4096 that have never been allowed on home pcs before. With the more TMU in the R600 ATI is trying to fix its texturing weakness it had in the R520/R580.

R580 and R520 never was weak in texturing. Efficency comes to mind, and at the very least these 16 decoupled TMUs were a match for G7x 24 TMUs. TMUs too have efficency and can vary in performance.

QFT, I was wondering where he was getting that from. ATi's texturing was in no way inferior to Nvidia's for the R5*** series.

 

[Cooler]

Originally posted by: josh6079

Originally posted by: Cookie Monster

Originally posted by: Cooler

Originally posted by: jim1976

Originally posted by: josh6079
My previous example was just an analogy. All I'm saying is that until we know what algorithms the architecture uses it is pointless to debate about the necessity for such bandwidth.

I really hope I'm wrong and it turns out to be a 512bit controller with 1GB ram and on top of that it will be fully utilised. I just don't see it possible..

It may allow for larger texture sizes greater than 4096x4096 that have never been allowed on home pcs before. With the more TMU in the R600 ATI is trying to fix its texturing weakness it had in the R520/R580.

R580 and R520 never was weak in texturing. Efficency comes to mind, and at the very least these 16 decoupled TMUs were a match for G7x 24 TMUs. TMUs too have efficency and can vary in performance.

QFT, I was wondering where he was getting that from. ATi's texturing was in no way inferior to Nvidia's for the R5*** series.

I never said the quality of textures was less then the G70. R520 had fewer units which meant they needed higher clock speeds. With more TMU the could decrease the clock speeds for better yields.

 

[thilanliyan]

Originally posted by: Cooler
I never said the quality of textures was less then the G70. R520 had fewer units which meant they needed higher clock speeds. With more TMU the could decrease the clock speeds for better yields.

Weren't the stock clocks for the 7900GTX at 650MHz and for the X1900XTX at 650MHz also?? So with fewer TMUs at the same speed ATI was getting better performance in most games (other than usually OpenGL games) so they didn't really need higher clocks either to get better performance, which means they were definitely more efficient than the G70/G71 counterparts.

 

[Cookie Monster]

Originally posted by: thilan29

Originally posted by: Cooler
I never said the quality of textures was less then the G70. R520 had fewer units which meant they needed higher clock speeds. With more TMU the could decrease the clock speeds for better yields.

Weren't the stock clocks for the 7900GTX at 650MHz and for the X1900XTX at 650MHz also?? So with fewer TMUs at the same speed ATI was getting better performance in most games (other than usually OpenGL games) so they didn't really need higher clocks either to get better performance, which means they were definitely more efficient than the G70/G71 counterparts.

Firstly they were decoupled compared to the G7x TMUs. It was a more efficent design. But even then the G7x or more precisely NV4x chips still were competitive.

 

[DeathReborn]

Perhaps to silence critics it needs the 1GB to put the CCC on VRAM so people can't call it "bloated".

512bit 1GB GDDR4 @ 2.4GHz does provide a hell of a lot of bandwidth but I question if the GPU can push pixels fast enough to make use of it. 64 vec4 shaders just doesn't quite sound powerful enough to use that bandwidth.

On the downside that's a lot of expensive GDDR & availability might be an issue as could price be. I suspect the R600 Core is cheaper to manufacture but the memory offsets than by some margin.

Last point, didn't the A1 spin still have problems running higher than 4xAA? Hopefully that isn't going to be a poblem with the A2 spin.

 

[Warren21]

Originally posted by: DeathReborn
Perhaps to silence critics it needs the 1GB to put the CCC on VRAM so people can't call it "bloated".

512bit 1GB GDDR4 @ 2.4GHz does provide a hell of a lot of bandwidth but I question if the GPU can push pixels fast enough to make use of it. 64 vec4 shaders just doesn't quite sound powerful enough to use that bandwidth.

On the downside that's a lot of expensive GDDR & availability might be an issue as could price be. I suspect the R600 Core is cheaper to manufacture but the memory offsets than by some margin.

Last point, didn't the A1 spin still have problems running higher than 4xAA? Hopefully that isn't going to be a poblem with the A2 spin.

Unless you can produce the artile that said so, I do believe it was A0 spin that had the problems only.

 

[rgreen83]

Originally posted by: Warren21

Originally posted by: DeathReborn
Perhaps to silence critics it needs the 1GB to put the CCC on VRAM so people can't call it "bloated".

512bit 1GB GDDR4 @ 2.4GHz does provide a hell of a lot of bandwidth but I question if the GPU can push pixels fast enough to make use of it. 64 vec4 shaders just doesn't quite sound powerful enough to use that bandwidth.

On the downside that's a lot of expensive GDDR & availability might be an issue as could price be. I suspect the R600 Core is cheaper to manufacture but the memory offsets than by some margin.

Last point, didn't the A1 spin still have problems running higher than 4xAA? Hopefully that isn't going to be a poblem with the A2 spin.

Unless you can produce the artile that said so, I do believe it was A0 spin that had the problems only.

Thats how i remember it also, A0 had those issues, fixed in A1 which is being tested and qualified and A2 will go on final boards.

 

[DeathReborn]

Originally posted by: Warren21

Originally posted by: DeathReborn
Perhaps to silence critics it needs the 1GB to put the CCC on VRAM so people can't call it "bloated".

512bit 1GB GDDR4 @ 2.4GHz does provide a hell of a lot of bandwidth but I question if the GPU can push pixels fast enough to make use of it. 64 vec4 shaders just doesn't quite sound powerful enough to use that bandwidth.

On the downside that's a lot of expensive GDDR & availability might be an issue as could price be. I suspect the R600 Core is cheaper to manufacture but the memory offsets than by some margin.

Last point, didn't the A1 spin still have problems running higher than 4xAA? Hopefully that isn't going to be a poblem with the A2 spin.

Unless you can produce the artile that said so, I do believe it was A0 spin that had the problems only.

CookieMonster mentioned it on page 6 of this thread. It's been talked about on B3D apparently.

 

[Cookie Monster]

A0 never even worked. A1 had those problems.

 

[sethk]

I think the R600 will be as radical a departure from the R5xx as the G80 is from the G7x. It won't be a total departure from the Xenos though, just adjusted to provide proper DX10 and DX9 performance, as opposed to the custom APIs (still closely related to DX) that the 360 uses.

Nvidia chose to go with scalar pipelines in a way that I think we can safely say ATI will not, so in that sense G80 is a bigger departure, but in the end it all comes down to actual performance, and in that sense I think it will be very surprising if R600 does not take the crown at release, even if ATI has to release a card that is much more expensive to make (and how can it not be with that pin count). Maybe the space savings from the 80nm process will partially offset that cost.