G70 Vs. R520 All things being equal

[wonkyturk]

Does the number pipelines stay proportional to the number of ROPs on the G70 or is always 16? If the the number of ROPs didn't decrease with the pipelines its an unfair comparision.

 

[Munky]

Originally posted by: CaiNaM

Originally posted by: munky

Originally posted by: CaiNaM
ya.. sure do.

but strickly from a technical aspect i still don't see anything useful here, other than perhaps it shows the r3xx architecture (and yes, other than the revamped memory architecture and some other tweaks, it's still more r3xx core than not) has pretty much reached it's limits, which explains why ati will go a completely different architecture here very shortly. course, we didn't really need this comparison to tell us that..

kind of like intel -- they could only go so far with minor core changes and ramping clockspeeds until they finally hit that "wall" with their technology...

Did you read anything about the r520 archtecture? Obviously not, because the g70 is alot more similar to the nv40 than the r520 is to the r300. If you really want to claim it's the same, then read this http://www.beyond3d.com/reviews/ati/r520/ and tell me if you stil think it's the same thing. In fact, the r520 will be the basis of the r580, which will emphasize shader power, and not just slapping on more pipes. Because they separated the texture units from the pixel shaders, it breaks the whole concept of a "pipe", and is in fact half-way there on the road to unified shaders.

lol.. did you?

i'll stick with what i said: r520 carries over most of the design principles of r420, and r300 before it. ati did revamp the memory architecture (as i acknowledged), however the vertex shaders and pixel shaders all follow the principles of r420 and earlier (the obviously had to extend it's capabilites to (almost fully) support sm3.

while the updated memory architecture did increase some of the efficiency (and granted, this is a bit of an oversimplification as there are other areas ati has attempted to optimize with varying degress of success, but that's beyond the scope of this dicussion), it was nowhere near enough of a design change to keep ati competetive. moving to 90nm was a must so ati could ramp up the clockspeeds (way beyond r420's) in order to keep pace.

so yea, all in all r520 is an evolution of the desing pricnciples of r420 and r300 before that. it may have been intentional (to smooth the move to 90nm) or they may have been forced to as they simply weren't ready to apply other techniques.. either way, more radical changes are in the forefrong as they just can't keep looking at raising clockrates (as they have the past 3 generations) to raise performance.

The basic ALU configuration in each shader stayed the same, but that's a big oversimplification of the overall archtecture changes. I'll stick with the theory that Ati will not be changing much for the r580, they'll just have 3 pixel shaders for every texture unit, like the rv530. The difference is only that it'll be a wider design, so instead of 12 shaders and 4 TMU's, the r580 will have 48 shaders and 16 TMU's, staying with the 3:1 ratio.

 

[Wreckage]

Originally posted by: classy
Comparisons like this are so stupid. They have absolutely ZERO relevance and offer ZERO insight.

They show a fair comparison of the two designs. Same clock and same number of pipes.

I don't understand why that would upset you?

 

[BenSkywalker]

What's particularly impressive is the 7800 uses a similar number of transistors to the X1800 so when comparing 16 pipes to 16 pipes the 7800 is using quite a few less (as it doesn't need 8 pipes worth of transistors) yet is still faster clock for clock.

This is a very major point. ATi's design is pushing the limits of what they have been able to achieve on the 90nm fab process and they can't compete against a 130nm part with a large section of its die disabled at comparable clocks. This has no real impact at all on the parts that are available now but I think it demonstrates quite clearly why ATi is abandoning the R3x0 core in its entirety with the R600 parts. What is interesting to note is that nVidia is also dropping their NV4x parts when they move to the NV50(G80)- it will be interesting to see how they stack up.

BTW- As stated before- this has no impact on the final products, this is more of interest to those that are interested in architectural choices and how they impact final products- end results are all that matters.

 

[CaiNaM]

Originally posted by: munky
The basic ALU configuration in each shader stayed the same, but that's a big oversimplification of the overall archtecture changes. I'll stick with the theory that Ati will not be changing much for the r580, they'll just have 3 pixel shaders for every texture unit, like the rv530. The difference is only that it'll be a wider design, so instead of 12 shaders and 4 TMU's, the r580 will have 48 shaders and 16 TMU's, staying with the 3:1 ratio.

the overall architectural changes are not that significant.

i mean.. perhaps if you're talking about changing anything or adding anything to a chip is a "major" change (and i suppose to some degree it is -- not like you can just bolt it on. depends on perspective), but again, the design and concept is pretty much r300. the "thinking" is the same.

the one true exception to this is the mem. architecture, which they actually tried to do differently.

and as long as we're in this thread, i would find it interesting if - all things being equal (core/mem speed) the x1800 is really any faster than an x850xtpe. i wouldn't thinks there would be much difference as IT'S BASICALLY THE SAME ARCHITECTURE

who knows, maybe DH can throw in an r420 and prove me wrong here. if i didn't think an xl is a poor value at this point i'd get one myself and do it. perhaps rollo can do it as he seems to have an umlimited budget as well as time to do extensive tests!

personally that would be a true tech comparison, rather than an nv vs ati as is the case now.

 

[Pete]

Originally posted by: wonkyturk
Does the number pipelines stay proportional to the number of ROPs on the G70 or is always 16? If the the number of ROPs didn't decrease with the pipelines its an unfair comparision.

Yeah, that's another thing to consider (G70 has more pipes than ROPs in its available configs), but I don't think you can disable ROPs, so they're stuck at 16. More unfair is that I don't think the 7800 was clocked at 450, but possibly at 459 for the pixel pipes and 490 for the vertex shaders, as G70 has clock plateaus. Unwinder discovered, or at least confirmed, this when everyone was searching for the G70's missing quads back when it was released.

Originally posted by: BenSkywalker

What's particularly impressive is the 7800 uses a similar number of transistors to the X1800 so when comparing 16 pipes to 16 pipes the 7800 is using quite a few less (as it doesn't need 8 pipes worth of transistors) yet is still faster clock for clock.

This is a very major point. ATi's design is pushing the limits of what they have been able to achieve on the 90nm fab process and they can't compete against a 130nm part with a large section of its die disabled at comparable clocks. This has no real impact at all on the parts that are available now[...]

I think ATI went to all this trouble in preparation for R580 and also RV530. They wanted to separate the pixel shader units from as much as possible (in this case, the texture units and the ROPs) so as to reduce the transistor cost (well, probably die size [scroll up], ultimately) of adding "pipes" (in the sense that matters most with future games: shaders). So, while G70 is close in transistor count to R520 but with 8 more pixel pipes now, I don't think nV will be able to yield a 48 pipe card with G70's pixel pipe architecture to compete with R580's 48 pixel shader units. Of course, they probably won't need to, what with their higher-IPC pipes and the extra TMUs (one per pipe, while R580 probably remains with 16). So we'll probably see nV field either a 550+MHz, 24-pipe G7x or a lower-clocked, 32-pipe G7x against ATI's probably 600+MHz, 48-pixel-shader-but-16-TMU R580. That's where things'll get interesting in terms of performance in future titles.

I have to imagine ATI will field something in between RV530 and R580, either a unique GPU (16-24PS, 8TMU, 4-8ROP) or a cut-down R580, both considering the gulf b/w RV515 and R520 (or RV530 and R580) and a potential 16PS, 8ROP nV part (nested b/w 8PS and 24PS parts). Didn't we hear that both ATI and nV will field a wider range of GPUs in the future? Maybe we'll see the R520 continued, but as a performance midrange ($300-400) part, when R580 assumes the enthusiast ($450+) mantle?

 

[classy]

Originally posted by: Wreckage

Originally posted by: classy
Comparisons like this are so stupid. They have absolutely ZERO relevance and offer ZERO insight.

They show a fair comparison of the two designs. Same clock and same number of pipes.

I don't understand why that would upset you?

No one is upset, its stupid. Lets see I am going to take a Vette, close off 4 of its 8 cylinders, then compare it to a SRT4. That sounds retarded doesn't it? If its built different, I don't see how turning something in one and then making a comparison is somehow equal? And even worse come up with a so-called conclusion. If the architectures shared similarties it would be of some interest, but they don't.

 

[DidlySquat]

what about the 512 MB ram vs 256 MB ? did they forget about that ??? i'm sure that's a major factor in the high-res, high-detail, high AA/AF tests. When the 512 MB 7800 GTX comes out it's going to be game over for ATI.

Edit: I apologize if anyone already mentioned that fact as I didn't read the whole thread

 

[BenSkywalker]

They wanted to separate the pixel shader units from as much as possible (in this case, the texture units and the ROPs) so as to reduce the transistor cost of adding "pipes" (in the sense that matters most with future games: shaders).

nV has already decoupled their ALUs from ROPs though, and did so with considerably less transistor useage.

So, while G70 is close in transistor count to R520 but with 8 more pixel pipes now, I don't think nV will be able to yield a 48 pipe card with G70's pixel pipe architecture to compete with R580's 48 pixel shader units.

But will ATi? If the supposition is that ATi will be moving over to their more simplistic ALUs with the R580 then it will be interesting to see what levels of performance trade off they end up with(in terms of the more powerful/complex ALUs v more quantitive/simple ALUs). Either way, given the edge that they are pushing with the 1800xt I don't see ATi bumping the transistor count too much without a serious downclocking of the part. In the end that may be a worthwhile tradeoff, not saying it won't, but it doesn't seem too likely to me that they are going to be able to approach 400m transistor part at ~600MHZ using the 90nm process.

I have to imagine ATI will field something in between RV530 and R580, either a unique GPU (16-24PS, 8TMU, 4-8ROP) or a cut-down R580, both considering the gulf b/w RV515 and R520 (or RV530 and R580) and a potential 16PS, 8ROP nV part (nested b/w 8PS and 24PS parts).

I think a lot of what direction ATi ends up taking depends on nV. When and for what chip are they ramping 90nm production? We know the G71 is supposed to be moving over, but what about the higher tier parts? nV is under no pressure now to do anything major(they can ship out a 512MB GTX clocked @500MHZ and be quite comfortable) so they can approach the process transition much as ATi did during the R3x0 days.

We know that the RSX is aiming for 32 ALUs, 16 ROPs and a clockspeed @550MHZ on 90nm- for a mass market consumer device, so one would assume that nV will be packing something more powerful in raw specs then that by next spring(as they have indicated). If they do, the 7800GT would be moved to a third or fourth tier card, much as the 6800GT is now- which would likely move the pricing in to the sub $300 range(as the 6800GT is entering now). I'm having trouble seeing the current(R52x) architecture being able to compete with that on the 90nm process, and we know 65nm is a long way off. While the yields are apparenty quite solid for high end parts, they don't seem to be of the caliber that would allow them to move the 1800xl down in to the segment they need to. I'm sure the situation will change as obviously right now ATi has a gaping x1800Pro sized hole in their lineup at the high end(or perhaps it will end up being the x1900Pro). It will be interesting to see, but it seems that moving forward nV has a decided edge at this point for the current generation of products and their revisions. I'm not thrilled with this as I really would love to see a great deal on a high end part with proper AF, but I'm not seeing that as being viable any time soon.

 

[Pete]

I didn't hear rumors of simplified ALUs, but rather the same (simple compared to G70's) ALUs as in R520 (and basically since R300). You're right, it may well be two similarly-clocked parts from ATI and nV at 90nm, albeit with different ALU and TMU amounts.

They can be comfortable with an upclocked G70, but if their yields are similar to ATI's with R520, they're getting fewer per wafer (assuming these pics are to scale: [l=scroll up]http://www.digit-life.com/articles2/video/r520-part2.html#p4[l]), so they'll want to move to 90nm eventually.

I thought RSX was still 24 pipes, essentially G70 @ 90nm with some mem tweaks to accomodate Cell access?

I agree that nV is better positioned ATM, at least in terms of high-end pricing.

 

[crazydingo]

Originally posted by: Pete
I thought RSX was still 24 pipes, essentially G70 @ 90nm with some mem tweaks to accomodate Cell access?

I agree that nV is better positioned ATM, at least in terms of high-end pricing.

Probably. Especially when you consider their 302 million transistor spec, it looks like an off the shelf G70 with added features (Turbocache) and a different interface (XDR?).

 

[Wreckage]

Originally posted by: classy

Originally posted by: Wreckage

Originally posted by: classy
Comparisons like this are so stupid. They have absolutely ZERO relevance and offer ZERO insight.

They show a fair comparison of the two designs. Same clock and same number of pipes.

I don't understand why that would upset you?

No one is upset, its stupid. Lets see I am going to take a Vette, close off 4 of its 8 cylinders, then compare it to a SRT4. That sounds retarded doesn't it? If its built different, I don't see how turning something in one and then making a comparison is somehow equal? And even worse come up with a so-called conclusion. If the architectures shared similarties it would be of some interest, but they don't.

It's more like someone with a cavalier telling someone with a corvette "the only reason the corvette is better is the extra 4 cylinders"

 

[ArchAngel777]

Originally posted by: Wreckage

Originally posted by: classy

Originally posted by: Wreckage

Originally posted by: classy
Comparisons like this are so stupid. They have absolutely ZERO relevance and offer ZERO insight.

They show a fair comparison of the two designs. Same clock and same number of pipes.

I don't understand why that would upset you?

No one is upset, its stupid. Lets see I am going to take a Vette, close off 4 of its 8 cylinders, then compare it to a SRT4. That sounds retarded doesn't it? If its built different, I don't see how turning something in one and then making a comparison is somehow equal? And even worse come up with a so-called conclusion. If the architectures shared similarties it would be of some interest, but they don't.

It's more like someone with a cavalier telling someone with a corvette "the only reason the corvette is better is the extra 4 cylinders"

QFT

 

[Keysplayr]

Originally posted by: classy

Originally posted by: Wreckage

Originally posted by: classy
Comparisons like this are so stupid. They have absolutely ZERO relevance and offer ZERO insight.

They show a fair comparison of the two designs. Same clock and same number of pipes.

I don't understand why that would upset you?

No one is upset, its stupid. Lets see I am going to take a Vette, close off 4 of its 8 cylinders, then compare it to a SRT4. That sounds retarded doesn't it? If its built different, I don't see how turning something in one and then making a comparison is somehow equal? And even worse come up with a so-called conclusion. If the architectures shared similarties it would be of some interest, but they don't.

Classy, you're going waaaaay overboard to get your simple point across that you think it's stupid. We heard you, understand it, and grasped it. Now that you have said your piece and know that we acknowledge you, are you satisfied? There are other people in here who do not think it's stupid. Can you deal with that?

 

[classy]

Originally posted by: keysplayr2003

Originally posted by: classy

Originally posted by: Wreckage

Originally posted by: classy
Comparisons like this are so stupid. They have absolutely ZERO relevance and offer ZERO insight.

They show a fair comparison of the two designs. Same clock and same number of pipes.

I don't understand why that would upset you?

No one is upset, its stupid. Lets see I am going to take a Vette, close off 4 of its 8 cylinders, then compare it to a SRT4. That sounds retarded doesn't it? If its built different, I don't see how turning something in one and then making a comparison is somehow equal? And even worse come up with a so-called conclusion. If the architectures shared similarties it would be of some interest, but they don't.

Classy, you're going waaaaay overboard to get your simple point across that you think it's stupid. We heard you, understand it, and grasped it. Now that you have said your piece and know that we acknowledge you, are you satisfied? There are other people in here who do not think it's stupid. Can you deal with that?

Its not a matter of dealing with it. I am just stating the obvious fact that comparisons like these truly offer no insight. Why? Because of the fact they are so different how does one determine any valid outcome of any tests or comparisons? Its similar to Intel and AMD, the only true valid comparisons are the results produced. Thats all I'm saying. And I won't say its stupid anymore, ok? But it is irelevant.

 

[ArchAngel777]

Originally posted by: classy

Originally posted by: keysplayr2003

Originally posted by: classy

Originally posted by: Wreckage

Originally posted by: classy
Comparisons like this are so stupid. They have absolutely ZERO relevance and offer ZERO insight.

They show a fair comparison of the two designs. Same clock and same number of pipes.

I don't understand why that would upset you?

No one is upset, its stupid. Lets see I am going to take a Vette, close off 4 of its 8 cylinders, then compare it to a SRT4. That sounds retarded doesn't it? If its built different, I don't see how turning something in one and then making a comparison is somehow equal? And even worse come up with a so-called conclusion. If the architectures shared similarties it would be of some interest, but they don't.

Classy, you're going waaaaay overboard to get your simple point across that you think it's stupid. We heard you, understand it, and grasped it. Now that you have said your piece and know that we acknowledge you, are you satisfied? There are other people in here who do not think it's stupid. Can you deal with that?

Its not a matter of dealing with it. I am just stating the obvious fact that comparisons like these truly offer no insight. Why? Because of the fact they are so different how does one determine any valid outcome of any tests or comparisons? Its similar to Intel and AMD, the only true valid comparisons are the results produced. Thats all I'm saying. And I won't say its stupid anymore, ok? But it is irelevant.

If it was an ovious fact, why did you have to tell us a few times? Why did you have to tell us even once?

 

[Keysplayr]

Originally posted by: classy

Its not a matter of dealing with it. I am just stating the obvious fact that comparisons like these truly offer no insight. Why? Because of the fact they are so different how does one determine any valid outcome of any tests or comparisons? Its similar to Intel and AMD, the only true valid comparisons are the results produced. Thats all I'm saying. And I won't say its stupid anymore, ok? But it is irelevant.

No, you are stating an obvious opinion. Not an obvious fact. And that's a fact. Leave it at that. I understand the whole INtel AMD thing. If you ran a prescott at the same MHz as an Athlon 64 3500 (2GHz) It would be fish paste. (SpongeBob SquarePants). One of the biggest arguments in the CPU forum is about efficiency and AMD abundance and Intel's lack of. My Pentium M notebook (1.73GHz Dothan) blows away my P4 3.0E at many levels.

Some people are insisting that the R520 doesn't have pipes at all, but instead has a sort of hybrid pipeline arch halfway to a unified arch. W....T....F..... goes through peoples minds in lengendary.

Anyway, I liked the article and found it interesting. I emailed Mr. Davidson at Driver Heaven and he responded within 30 minutes. I asked if he would be kind enough to add more game benches to his article. He said he was glad we enjoyed the article but was very swamped at the moment. But said he would try and squeeze some in if time permits him to.

So we "might" see some more added.

 

[Pete]

Does no one in the AT forums know how to quote without referncing the entire frickin thread? Learn to use the DEL button, ppl. There's a whole key on your keyboard dedicated to it. And there's no point in quoting someone when you reply right below his post!

Edit: Forgot a to make it playful.

And yes, keys. Thanks. I'll put the weapon of minor destruction down, now.

 

[Keysplayr]

That better Pete?

 

[ArchAngel777]

Originally posted by: Pete
Does no one in the AT forums know how to quote without referncing the entire frickin thread? Learn to use the DEL button, ppl. There's a whole key on your keyboard dedicated to it. And there's no point in quoting someone when you reply right below his post!

Edit: Forgot a to make it playful.

And yes, keys. Thanks. I'll put the weapon of minor destruction down, now.

Here I thought Pete was the only calm person on this board. I guess my theory is right, everyone gets peeved...

 

[Gamingphreek]

There's a whole key

Actually there are two keys.

The "." key with numlock off, and the "Delete" key. (At least for us standard english users).

Just to emphasize the point lol.

-Kevin

 

[Ackmed]

What a useless "article".

 

[Keysplayr]

Originally posted by: Ackmed
What a useless "article".

So it's a lot like your post here?

 

[christoph83]

Originally posted by: keysplayr2003
So it's a lot like your post here?

I'm with you on this one keysplayr2003. I think it's a good article. It doesn't neccesarily show which card is better, but more or less which one is running at a better efficiency.

In the end though it's like the AMD vs intel comparison. If intel had a 6ghz chip, it wouldn't really matter how slow it is per clock.

 

[Ackmed]

Except not. There are far too many things left unchecked, or not mentioned.

For a much better discussion about this article; http://www.beyond3d.com/forum/showthread.php?t=24465

Read thru that, and ask yourself if you still think the article is worth anything.