How can Nvidia be ahead of ATI but 360 GPU is on par with RSX?

[seismologist]

Why is NVidia so far ahead of ATI in the PC space yet the 360 (ATI) is launching this fall with a GPU that's supposed to be as good as the one in PS3 (PS3)?

It seems like since Nvidia is already ahead in their GPU tech. So If both consoles are using the latest technology at their respective launches then whatever in the PS3 should be at least one step ahead of the 360.

Can someone give me a quick summary of what's going on?

 

[dukmahsik]

Why is NVidia so far ahead of ATI in the PC space yet the 360 (ATI) is launching this fall with a GPU that's supposed to be as good as the one in PS3 (PS3)?

It seems like since Nvidia is already ahead in their GPU tech. So If both consoles are using the latest technology at their respective launches then whatever in the PS3 should be at least one step ahead of the 360.

Can someone give me a quick summary of what's going on?

who said nvidia is ahead in gpu tech? this is so wrong on so many accounts. rsx = current gen gfx as 7800 is already out, xenos = next gen gfx as r600 isn't out yet.

 

[kyleb]

Yeah, it makes things confusing when you start with a a false assumption.

 

[Acert93]

Why is NVidia so far ahead of ATI in the PC space yet the 360 (ATI) is launching this fall with a GPU that's supposed to be as good as the one in PS3 (PS3)?


Can someone give me a quick summary of what's going on?

R520 has been around since late Winter/early Spring. The first go around, as usual, required some tweaking. The second time it came back it did not seem right, and after some digging it turned out there was a soft ground issue that required yet ANOTHER spin--but pretty late in the deployment timeframe. This led to a delay in the R520. If all had gone well R520 could have been released 3 months ago.

So not a tech issue, but a "oh crap!" one. That is part of the risk of going on a new edge process on a HUGE chip. This is why GPU IHVs usually do their midrange on the new process first.

Xenos had working silicone last November. They probably had a respin etc and have been waiting for the 90nm process at TMC (or wherever they are fabbing it) to mature before production.

And it seems the success of Xenos at 90nm was what initially prompted ATI to go with 90nm for R520. Interestingly not all the R5xx parts were affected by the soft ground issue. Xenos and R515, I think, are both unaffected. Of course Xenos is at a modest 500MHz and has some of its logic on a daughter die (and if R520 is ~300M transistors like RSX, Xenos is a about 50M smaller).

That is about it... at least from what I have read. As for being advanced, Xenos is a followup to the ideas found in the unreleases R400 and seems to have a lot in common with the DX10 API an ATI's upcoming R600. While not identical and designed for the console marketplace from the ground up (one advantage it has) it will seemingly share similar DNA as the R600.

Ps- Feature wise Xenos has a lot of fresh ideas and advancements. FP10, FP10/16 with MSAA, MEMEXPORT, hardware tesselation, unified shader array with a pretty nifty scheduler, eDRAM with logic designed with 4x AA in mind, 32bit percision throughout the shaders, CPU cache locking where the GPU can read from the CPU cache, etc... and a lot of overlap inbetween (e.g. the double pumped Z with the eDRAM, then using all the ALUs to work on geometry... this should make shadowing nice).

Basically Xenos is a new GPU, not an evolution of R300. While it will surely have weak points, it offers up a lot of new ways to deal with problems. It is the evolutionary step that ATI is offering for common GPU issues. R600 will continue on in this vien, but with solutions with the PC desktop market in mind (Xenos is VERY console focused in design).

 

[Titanio]

who said nvidia is ahead in gpu tech? this is so wrong on so many accounts. rsx = current gen gfx as 7800 is already out, xenos = next gen gfx as r600 isn't out yet.

Xenos != R600.

Similar architectures, but similar power? I should certainly hope not, if R600 is coming late next year.

One thing that would be interesting to compare is if their relative experience with SM3.0(+) architectures yields any differences. NVidia jumped first, RSX is of a second implementation of that, whereas Xenos is at least amongst ATi's first (if not its first). Quality of implementation may not be a major point, but there was an improvement between NVidia's first and second.

We'll probably never know though..

 

[overclocked]

Read the article Dave put up on Xenos, search the forums and im sure you will se why.
Soon ATI should be equal atlest in PC-space, they have lacked SM3 and had various delays and so on. But read around, google and you will learn the differences.

 

[seismologist]

Thanks Acert. I was under the impression that Xenos hadn't taped out yet since 360 Beta kits were using X800 GPU's. And also because R520 delays I thought they were having trouble keeping up with Nvidia.

 

[Titanio]

Thanks Acert. I was under the impression that Xenos hadn't taped out yet since 360 Beta kits were using X800 GPU's.

Was Xenos not in the Beta kits? I thought it was "feature complete", just underclocked? Also, I believe Xenos taped out in Q1.

 

[ERP]

Was Xenos not in the Beta kits? I thought it was "feature complete", just underclocked? Also, I believe Xenos taped out in Q1.

Xenos was in Beta kits.

X800 and 9800's were in Alpha kits.

 

[seismologist]

Ok I think I get it now. So Xenos was on the same timeline as Nvidia's 7800 release but it's designed with some console specific features.

The RSX just recently taped out, so it's either 6 months newer technology or a cost reduction, we dont have all of the details on this chip yet.

Does that sound about right?

 

[scooby_dooby]

Well not really.

Xenos wasn't designed with "some" console features, it was designed compeltely from day 1 as a console GPU. Also the features that they have implemented into Xenos are more in-line with the R600 they will be releasing than the R520 which was on the same timeline as the 7800.

As for the 6month wait for RSX, isn't that customizations to the PC card to make it more suitable for the console? I would'nt think it's necessarily "newer technology or a cost reduction" more along the lines of customization for the console.

 

[Acert93]

Ok I think I get it now. So Xenos was on the same timeline as Nvidia's 7800 release but it's designed with some console specific features.

Not just console specific, but relying heavily on ATI's next gen part. G70 is an evolution of the NV40; Xenos is not an evolution of R300/R420.

The RSX just recently taped out, so it's either 6 months newer technology or a cost reduction, we dont have all of the details on this chip yet.

All we know about RSX is that from the original press release that it is an adaptation of G70, on 90nm, and is being clocked at 550MHz. The few performance numbers Sony gave seem to line up with the idea of RSX being a higher clocked G70 (although there wre a couple misplaced numbers, but not by much).

As for 6 month newer technology, newer than what? G70 or Xenos? Xenos is on a totally different development timeline. Xenos is basically stealing ideas from a PC part/platform that wont be available until it seems late Fall 2006. I know this goes against common sense, but a newer chip is not always a more "advanced" chip, especially if its base platform is older. e.g. CELL is more advanced than an X86 processor in many ways, so a new x86 release wont automatically mean it is more advanced/powerful (relatively... a lot of factors to take into consideration).

Whatever the case may be, we know RSX is faster than G70. At least the shader logic is (memory is another discussion for another day... search feature on that!) RSX wont have unified shaders, eDRAM, hardware tesselation, etc. Whether RSX or Xenos is faster (or which is faster in which situations) is an issue of debate and conjecture and we wont have a good understanding of this for probably 3 years, and even then it will be difficult because they are embedded chips in a closed platform, meaning they are highly influenced by they environmental limitations.

Ps- Read Dave's Xenos article. Firingsquad and HardOCP also have good articles. But read Daves. It answers almost all your questions. Or should. Wont answer the "Which is faster?" question, but then again that is almost irrelevant.

 

[Shifty Geezer]

Calling G70 'this-gen' and Xenos 'next-gen' seems a bit far fetched to me. If G70 is 'this-gen', what's ATi's 'this-gen' component? Both GPU's were in development around the same time, and are being released around the same time. Hence both are 'next-gen' GPU's, just with different ways of doing things. Although, I don't know how anyone can really class a generation of GPU's short of the designers' nomenclature. More important than it's generation is how well it actually performs at it's job

 

[Titanio]

RSX wont have unified shaders, eDRAM, hardware tesselation, etc..

These things are rattled off as disadvantages/advantages..it doesn't have unified shaders, but it does have dedicated shaders. It doesn't have edram, but it does have more computational logic. Etc. etc. Everything is a tradeoff. These things are not unequivocally or certainly better. RSX and Xenos are different solutions to the same problems, but with some differing priorities. Unified shading, for example, seems primarily a solution that would be of most benefit in the PC space actually (impossible to code specifically for one card, so let the card mould your code around itself). Xenos adopts upcoming features in the PC space, but the question is if it'd be "better" than a newer chip of a more traditional architecture. I think it's difficult to say that, especially when some of the main things it introduces aren't necessarily absolute wins with no tradeoffs vs the older solution, particularly in the context of the closed systems they're being put into.

edit - also, re. FP10 - RSX supports FP10, FP16, FP32 etc.

 

[Acert93]

Calling G70 'this-gen' and Xenos 'next-gen' seems a bit far fetched to me. If G70 is 'this-gen', what's ATi's 'this-gen' component?

R520.

Both GPU's were in development around the same time, and are being released around the same time. Hence both are 'next-gen' GPU's, just with different ways of doing things. Although, I don't know how anyone can really class a generation of GPU's short of the designers' nomenclature. More important than it's generation is how well it actually performs at it's job

There is a difference between features and performance.

To put it simply:

• G70 is designed with DX9/SM3.0 in mind with the desktop as the target platform.
• Xenos is designed with many DX10 elements in mind with a console as the target platform.

I listed a good number of features above (unified shaders array, MEMEXPORT, eDRAM, FP10, fully orthogonal ROPs in regards to FP10/16 and MSAA, hardware tesselation, etc) that are more advanced features. That does not mean every single one improves performance or IQ, but they are features not found on DX9 hardware (or even desktop parts in some circumstances).

Overlooking the differences is about equivlant as looking at Xenon/CELL, noting they both are PPC based, multicore, same frequency and developed on a similar time frame, and shrug. They may be comparable, and even perform similarly, but the architectures are VASTLY different. And in certain circumstances one design is advantageous over the other.

On the GPU side of the equation there is no question that Xenos is much more advanced than G70 in regards to features. Performance, of course, is the unknown.

No one is saying RSX is not next gen; but from a design perspective--and its target (DX9)--it does incorperate "older" technology. It is not a clean slate design.

But before everyone gets defensive, I will paraphrase Dave from a long time ago: Xenos is high risk, high reward. It is a TOTALLY new design. It could be extremely effecient and powerful, or it may end up having significant issues that were unaccounted for. That is what happens with first generation clean slate designs. CELL runs the same risk. Obviously these are calculated risks (e.g. Xenos is not a completely new design in that it is following the R400, so there was a test case) but risks none the less.

Sometimes the "tried and true" method is not only safer, but also more effecient/effective. So noting that RSX is from the G70 line (which NV claims it is) and Xenos shares more in common with ATI's unreleased R600 series (which cannot be released without an API! So there is more to the issue when considering the "Whys and Hows") is not some diss on RSX.

But there is no question Xenos has some features RSX does not.

 

[Acert93]

These things are rattled off as disadvantages/advantages..

No, they are features. Features that Xenos has that are part from future GPU lines. You cannot sit here and tell me hardware tesselation is not an advanced feature, or being able to perform floating point blending with MSAA is not an advanced feature.

it doesn't have unified shaders, but it does have dedicated shaders. It doesn't have edram, but it does have more computational logic.

Actually the computational logic is much closer than commonly presented. Typically people are tossing around the 232M vs. 300M numbers from Xenos and RSX, respectively. But Xenos has another 20M logic transistors on the eDRAM, and RSX, if a G70 adaptation, will have 20-25M transistors from Purevideo. ~250 vs. ~275 is a meager difference. To compare, R420 was 160M and NV40 was 222M. Not only is the total smaller for Xenos/RSX, but the percentage is significantly closer.

One could argue that more computational logic does not mean faster (especially given different designs), but that would undermind the, "RSX has more computational logic" point

Etc. etc. Everything is a tradeoff.

Of course. But not all tradeoffs are equal. NV did not employ SM3.0 in NV40, and then refine it with a lot of extra silicone in G70, because it was a waste--it has advantages. They did not put extra work in beefing up the ROPs for better HDR performance for no reason either.

Some tradeoffs are speed, some are for IQ, and some are for effeciency. G70 is designed with the limitations of DX9/PCs in mind.

Xenos was not. That does not mean it will be better, but ignoring this point is silly. The tradeoffs are not the same because the limitations are different.

These things are not unequivocally or certainly better. RSX and Xenos are different solutions to the same problems, but with some differing priorities. Unified shading, for example, seems primarily a solution that would be of most benefit in the PC space actually (impossible to code specifically for one card, so let the card mould your code around itself).

That is partly a myth; it has been answered before and certain people continue to push. During the rendering process in a scene there are events that are either PS intensive or VS intensive by their very nature. While on a console you can design your game assets to attempt to maximize, and properly balance, the rendering load, the actual rendering process on each scene has its own spikes and dips that cannot always be avoided. In these situations unified shaders are a huge advantage.

In the most extreme cases you will have 48 ALUs all doing vertex work. That is a significant advantage over having 10 VS doing the same task.

Dave and Jawed has explained this more than once, and I am sure that he will have to explain it again.

Xenos adopts upcoming features in the PC space, but the question is if it'd be "better" than a newer chip of a more traditional architecture.

You are argueing with a make believe person. You quote me to begin your discussion, but I never said it was better. I said it had more advanced features. I even proposed the idea that a traditional, proven, design may be "better".

Obviously 1st party developers are going to design to the strengths of a platform (at least at some point in 2 or 3 years). You have participated in a lot of threads about where "CELL will do this better than Xenon". You should understand that certain designs give certain advantages in specific situations. A good developer exploits this.

The difference, if there is one, is that GPUs have well known hurdles ahead of them. There is a clear evolution in the APIs, and the hardware to support these APIs, to resolve these issues.

G70 is designed with supporting DX9's solutions to these problems. G70 does not have DX10 features, which can be a solution to many problems, because DX10 is not the target platform.

DX10, which is around the corner, is the "next step" in solutions to these problems. Just like programmable shaders were a big jump in DX7, and have continued to evolve through DX8, DX8.1, and DX9, so the next step is DX10.

None of this is to say one chip is better blah blah blah. But it is also unfair to minimize the advances of a chip when it is able to use new techniques because it is not closely tied to an older API. To dismiss them as "tradeoffs" really is undervalueing the advances.

If that were the case we would still be using DX6 GPUs and they would still be on the market--because they are, after all, just tradeoffs. I mean who needs programmable shaders?

 

[Shifty Geezer]

Overlooking the differences is about equivlant as looking at Xenon/CELL, noting they both are PPC based, multicore, same frequency and developed on a similar time frame, and shrug. They may be comparable, and even perform similarly, but the architectures are VASTLY different. And in certain circumstances one design is advantageous over the other.

And would you call one of these processors 'next-gen' and the other 'current-gen'? I don't think so. I appreciate the differences of Xenos and advanced features, but as i say I don't think it can be called 'next-gen', at least not counting a generation as the next GPU after the previous one.

No one is saying RSX is not next gen

I beg to differ ...

who said nvidia is ahead in gpu tech? this is so wrong on so many accounts. rsx = current gen gfx as 7800 is already out, xenos = next gen gfx as r600 isn't out yet.

And it's primarily to this that i responded. If dukmahsik is right, RSX is a generation behind Xenos, which I would understand to be about half as powerful given the way GPU's progress. As there's no real definition of a generation in GPU terms (if Shader Model is considered ATi didn't have a generation advance in AGES!) it's not really a great way to talk about GPU's, specially seeing as the term is full of connotations on performance.

 

[blakjedi]

R520 is ATI's DX9/SM3.0 full featured desktop part. It is considered in the same generation as the G70 as it designed with contemporary performance to meet the SM3.0 API

ATI will probably release refreshes to this part between now and next fall when their "next generation" part, the R600 should launch.

RSX as far as we know is essentially an advanced "refresh" of the G70 based on the g70 itself.

Nvidia will release a part that will compete more definitively with the R600 next fall. That part will be their next generation part.

Xenos is based more on the tech that will be in the R600 than on the R520 (in fact some say R520 may have some elements of the Xenos in it!). Making it much more accurately a next gen part.

It doesnt get more simple than that. Dont let the impressive brute force of the RSX confuse you into think its more next gen than Xenos. Also next gen doesnt necessarily mean better but it does mean more advanced... whether the advances were done well is the question.

For example 6800 and G70 are clearly a generation beyond the X800 but the x800 series has compeitive performance if not feature set with those more advanced chips from Nvidia...

 

[Jawed]

It's simple: RSX is NV40 with some tweaks (symmetrical ALUs in each fragment pipe being the most obvious) and running faster on 90nm. Started back in 2002.

Xenos was started back in 2002 as well, but is about 2 NVidia generations ahead. NVidia won't have anything this advanced until after G80. Call it G90, if you like - 2007.

Jawed

 

[Acert93]

And would you call one of these processors 'next-gen' and the other 'current-gen'?

I would say Xenos has more in common with a more advanced API and has features to support such.

"Next Gen" is a very arbitrary word in certain contexts. Works well for consoles, or items competing in a set time frame on a designated platform. But G70 has a different API focus (DX9) and platform (PC) than Xenos (appearing to be more and more DX10 friendly, and console).

But that is still features; performance is a totally different ballpark. e.g. a Ti4200 (DX8.1) kicks the FX5200 (DX9) butt!

I beg to differ... And it's primarily to this that i responded.

I guess that is what I get when I skim certain members posts hahahaha It is his error though. The G70 is "next gen" in that it is brand new! But it is also a refresh of NV40 (no debates on that, it clearly is).

So yeah, my bad.

If dukmahsik is right, RSX is a generation behind Xenos, which I would understand to be about half as powerful given the way GPU's progress. As there's no real definition of a generation in GPU terms (if Shader Model is considered ATi didn't have a generation advance in AGES!) it's not really a great way to talk about GPU's, specially seeing as the term is full of connotations on performance.

Well, people cannot toss around comparisons like that as we know. Ti4200 and FX5200 and all!

GPUs can be measured in a couple ways.

Refresh (NV40=>G70; R300=>R360)
Evolution of a platform (R300=>R420; NV40=>G70, note the overlaps!)
API (DX7, DX8, DX9, etc.)
Architecture (R300 => R520 => R600; NV20 => NV30)

So, uh, yeah. Hard to say "this is a generation!" Generation of *what*?

As for performance, performance tends to double with GPUs ever 18-24 months. Now obviously two similarly sized chips, on different APIs (e.g. DX6 and DX7) will have different limitations. New APIs have brought new techniques like normal mapping, HDR, volumetric fog, programmable shaders, etc...

The best way I can describe Xenos and RSX is obviously similar performance class (class, key word... class can mean a lot of things; vague intentionally) due to their size and release time. Similar transistor counts for logic, similar frequencies, same process, etc.

In terms of feature set Xenos obviously is more specialized for the console market (e.g. eDRAM) and is tailored around a more advanced API. How important that will be is anyones guess and is conjecture at this point. To outright dismiss it is wrong.

But on the other hand we need to take it with a large grain of salt. Maybe that is hard to do for those who chomp at the bit at the performance wars (not you Shifty). But look at this gen. As different as the consoles are they all competed pretty evenly.

And on a closed box think of NV30 and R300. Sure, R300 kicked butt in PC space. It was way more DX9 friendly. But guess what, in the console world some of that does not matter. Sure, NV30 blew chunks at FP32 and had not FP24 support and had register issues. But a smart dev could work with that (it screamed in DX8.1). So customize your toolset and API to the GPU and selectively use FP32 where needed.

Same thing with the console processors (I really want to say "stupid console processors" because this same debate is sooo common). They each have limitations. Devs will work around that in most cases. Usually a strength can compensate for a weakness. And where not, well, it means the weakness in one game is a strength in another. Really, what is the difference between 10,000 physics objects and 5,000?

So what if one GPU cannot do a technique. Look at bloom--we have been faking HDR forever!

GPUs are a little more linear than CPUs, granted. There are known issues, and when the transistor realestate opens up implimenting new solutions can solve a problem more effeciently. This is why sometimes a new GPU is only a certain percentage faster at older games, but 3 or 4x fast at a new game. It is designed to speed up these new techniques and issues.

Not much need to beef up hardwired T&L when you have programmable VS. But on a console that can be overcome in many ways. And in the ways not, oh well. It means one less bell or whistle (or in the physics example less physics stuff).

Not sure how people can be so strongly for CELL and its advantages and yet dismiss that Xenos has advantages.

We knew MS and Sony had different paradigms from day one. Where the memory controllers are placed in the respective systems is a dead giveaway.

Personally, I quite enjoy the difference. They each have hurdles (how do we deal with so many cores effeciently?) and questions (what techniques will these features in Xenos help, if any????) But nothing to get bent out of shape over!