Can you calculate REAL WORLD Specs with any accuracy?

[BenQ]

When we talk specs we typically only talk the specs handed out by the company manufacturing the hardware, but these specs are theoretical maximums.

For example, maybe a next gen console will be listed as being able to push 500 million ( random number ) polys per second. But we know full well that at 500 miullion polys per second were talking about nothing more than wireframe models. So the question becomes, how many polys can it REALLY push in a real-time gaming application. ( It's the only thihng that matters )

The same scenario applies to the often debated Unified vs Standard shaders topic. * When I refer to Unified/Standard shaders I mean the Xenos, and RSX. *

The RSX does have higher theoretical maximums but as we all know, standard shaders are inefficient, but HOW inefficient? ( ATI claims 50-70% efficient ). Meanwhile Unified shaders have lower theoretical maximums, but are ( we are told ) far more efficient ( apprantly 95-100% efficient ), meaning that unified shaders can come much closer to actually achieving their theoretical maximums.

The whole issue becomes one of efficientcy. I cannot say for sure but Unified shaders COULD wind up being far more powerful in REAL WORLD applicaton despite a lower theoretical maximim.

These scenarios apply to petty much any spec, so is there anything we can do? We are here at this forum to debate hardware, so being able to accurately gauge REALworld performance is extremely important.

Do we simply have to wait for devs to tell us the kind of real world performance they are getting in relation to the game they're working on?

 

[Titanio]

Many of the issues raised here have been discussed a lot at one point or another. But..

The RSX does have higher theoretical maximums but as we all know, standard shaders are inefficient, but HOW inefficient? ( ATI claims 50-70% efficient ). Meanwhile Unified shaders have lower theoretical maximums, but are ( we are told ) far more efficient ( apprantly 95-100% efficient ), meaning that unified shaders can come much closer to actually achieving their theoretical maximums.

Utilisation is a better word here than efficiency. Xenos could have higher utilisation than RSX. On a per shader level, RSX could have higher efficiency. Also remember we're dealing with closed boxes, where utilisation can be higher generally than in, for example, PCs.

Do we simply have to wait for devs to tell us the kind of real world performance they are getting in relation to the game they're working on?

To a degree, yeah. We can only go so far on paper. Although some "paper contrasts/comparisons" are so distinct that they're likely not to be so different in real scenarios. And some are less solid, yeah.

 

[Guden Oden]

Meanwhile Unified shaders have lower theoretical maximums, but are ( we are told ) far more efficient ( apprantly 95-100% efficient ), meaning that unified shaders can come much closer to actually achieving their theoretical maximums.

That statement strikes me as rather misleading, particulary considering both nvidia and ATI agrees unified shaders have efficiency issues, due to the different needs of vertex and pixel processing. There's a reason behind nvidia's expressed reluctance to go the unified route, and not because they aren't capable...

Besides, there's really not any such thing as "real-world specs"... Just check out PC games for an example of that. By going by 3dmark 05 results, all current 3D accelerators perform rather abysmally. However, going by quake3 results, they do much better.

It all depends on how the software running on it is written. Some software will inherently fit a certain architecture better than another (often suspected with nvidia-stamped twistbp titles on internet forums ), and generally speaking, the more of the functions of a 3D accelerator a title utilizes, the slower it tends to run.

So getting a real-world performance metric isn't possible. One could establish a metric of some sort of a (preferably large) selection of titles and measure the average framerate on a set hardware and software configuration, but that'd be fairly pointless too...

 

[Bohdy]

I think that the next-gen architectures are just too complex (for me at least) to estimate complete real-world performance with any degree of accuracy. You can look at one aspect at a time, but the big picture would require much more than idle speculation.

The games will be the judge.

 

[BenQ]

That statement strikes me as rather misleading, particulary considering both nvidia and ATI agrees unified shaders have efficiency issues, due to the different needs of vertex and pixel processing. There's a reason behind nvidia's expressed reluctance to go the unified route, and not because they aren't capable...

I think you just have to look further into nVidi's doublespeak on the issue.

In one breath they claim to not see the benefits of unified shaders and cry about the complexities of the logic necessary for proper load balancing. In the next breath they say they will be implamenting Unified shaders into thier next gen GPU.

It's realy not hard to see what's going on here. The simple fact is that if Unified shaders weren't superior overall we would NOT be seeing both ATI and nVidia moving toward Unified shaders. ATI has simply beat nViia to the punch. We will see nVidia downplaying unified shaders UNTILL nVidia has their very own in a future GPU, THEN they will claim it was them who "got it right."

And where has ATI claimed that their unified shaders have "efficientcy issues"? Everything I have read suggests that the WHOLE POINT in moving towards Unified shaders is to increase efficientcy beyond what standard shaders are capable of, due to their very architecture, and what ATI is claiming is that they are damn near 100% efficient.

Besides, there's really not any such thing as "real-world specs"... Just check out PC games for an example of that. By going by 3dmark 05 results, all current 3D accelerators perform rather abysmally. However, going by quake3 results, they do much better.

It all depends on how the software running on it is written. Some software will inherently fit a certain architecture better than another (often suspected with nvidia-stamped twistbp titles on internet forums ), and generally speaking, the more of the functions of a 3D accelerator a title utilizes, the slower it tends to run.

So getting a real-world performance metric isn't possible. One could establish a metric of some sort of a (preferably large) selection of titles and measure the average framerate on a set hardware and software configuration, but that'd be fairly pointless too...

I guess I can't dissagree with any of that, but I can't accept that a ballpark figure for real world specs is impossible to know. Even a ROUGH ballpark figure would HAVE to be more accurate than theoretical maximums.

 

[Titanio]

I think you just have to look further into nVidi's doublespeak on the issue.

In one breath they claim to not see the benefits of unified shaders and cry about the complexities of the logic necessary for proper load balancing. In the next breath they say they will be implamenting Unified shaders into thier next gen GPU.

It's realy not hard to see what's going on here. The simple fact is that if Unified shaders weren't superior overall we would NOT be seeing both ATI and nVidia moving toward Unified shaders. ATI has simply beat nViia to the punch. We will see nVidia downplaying unified shaders UNTILL nVidia has their very own in a future GPU, THEN they will claim it was them who "got it right."

That's a valid interpretation, but not the only one.

"Getting it right" isn't just about doing something, it's also about timing. Was putting FP32 precision in NV3x getting it right? Not to draw the comparison - I'm not saying it was a mistake to go with unified shaders now -just that doing something first isn't always a good idea. Some things can be premature.

In other words, saying "we don't want it now, but we'll do it later" isn't necessarily contradictory.

Remember, also, that NVidia is focussed around the PC market at the moment. Sure, RSX is there, but it's a derivative of work in the PC area. Does ATi have a unified architecture in the PC space currently? Why not?

The question as it concerns consoles is a lot more complex. I mean, if the overhead of low balancing is acceptable and individual efficiency of the shaders does not drop too low, all else being equal, I'd easily take the unified architecture in an open system (PC) where code can rarely be targetted at one specific ratio of vertex : pixel shaders or one particular featureset. However, when all else is not equal, and you're dealing with a closed system, where there may be some drop in per-shader efficiency, things are a lot less clear IMO.

 

[BenQ]

That's a valid interpretation, but not the only one.

"Getting it right" isn't just about doing something, it's also about timing. Was putting FP32 precision in NV3x getting it right? Not to draw the comparison - I'm not saying it was a mistake to go with unified shaders now -just that doing something first isn't always a good idea. Some things can be premature.

In other words, saying "we don't want it now, but we'll do it later" isn't necessarily contradictory.

Remember, also, that NVidia is focussed around the PC market at the moment. Sure, RSX is there, but it's a derivative of work in the PC area. Does ATi have a unified architecture in the PC space currently? Why not?

The question as it concerns consoles is a lot more complex. I mean, if the overhead of low balancing is acceptable and individual efficiency of the shaders does not drop too low, all else being equal, I'd easily take the unified architecture in an open system (PC) where code can rarely be targetted at one specific ratio of vertex : pixel shaders or one particular featureset. However, when all else is not equal, and you're dealing with a closed system, where there may be some drop in per-shader efficiency, things are a lot less clear IMO.

I honestly don't see how timing has anything to do with this. Unified shaders are completely tranparent to develpers. They are simply the solution of the problem that Standard shaders have. How can there be a wrong time to increase efficiency?

The ONLY problem I can see is IF The Xenos's Unified shaders DON'T work as they claim. It wouldbe the "wrong time" only if they DIDN'T "get it right."

......But if ATI did "get it right" then NOW is the perfect timing.

 

[Wunderchu]

I honestly don't see how timing has anything to do with this. Unified shaders are completely tranparent to develpers. They are simply the solution of the problem that Standard shaders have. How can there be a wrong time to increase efficiency?

The ONLY problem I can see is IF The Xenos's Unified shaders DON'T work as they claim. It wouldbe the "wrong time" only if they DIDN'T "get it right."

......But if ATI did "get it right" then NOW is the perfect timing.

I agree

 

[Titanio]

I honestly don't see how timing has anything to do with this. Unified shaders are completely tranparent to develpers. They are simply the solution of the problem that Standard shaders have. How can there be a wrong time to increase efficiency?

The ONLY problem I can see is IF The Xenos's Unified shaders DON'T work as they claim. It wouldbe the "wrong time" only if they DIDN'T "get it right."

......But if ATI did "get it right" then NOW is the perfect timing.

"Not getting it right" can be due to timing. For example, it may be argued that SM3.0+ level pixel and vertex shaders are too distinct still to merge, and thus a drop in per-task efficiency would be unacceptable. If a unified architecture experienced such a drop, it could be attributed as much to timing as to "getting it right" i.e. if one waited for later shader models.

Unified shaders may also be right for one system, but not right at this time for another. This harks back to whether "all else is equal".

As far as the PC space is concerned, it's still even more complex. DirectX does not support unified shading, and when it does, IIRC, it may no longer insist on the shaders being merged on a hardware level. NVidia will have to embrace unified shading from a software perspective, but will they merge on the hardware level, or do so initially? Also, I'm not sure if either company will want to be the first to introduce them in the PC space..there's some speculation that existing and near-future PC games wouldn't work as well with unified shaders as traditional architectures. If that were the case, I think both companies would want to "go together" to avoid risking even a short term drop in (apparent) performance relative to the competition.

 

[Shifty Geezer]

One question that concerns me is if ATi have got it right, why isn't their next PC card using Unified Shaders? Surely if they have that realworld advantage they'd use it?

"Hey, chief! These Unified shaders outperform standard shader by as much as 25% as they're more efficient, and we've got the technology when our rivals haven't."
"That's fantastic. Now let's not use it."

:?

I don't think the advantages over standard shaders can be that pronounced, both because of this and because nVidia have said they evaluate all the options and find the current system still works best. I question the efficiency of maybe the US in pixel shading. We've heard vertex shaders and pixel shaders are similar in operation which is why they can be unified. We also hear Cell is great at vertex work but not pixel work. That means that the difference between vertex and pixel, though slight, must have quite a performance difference. And as such an optimised pixel shader would, through the principle of specialsed>generalised, outperform a US running the same pixel work. Or it's just the overhead of the US logic takes away silicon from actual pipelines, so the increase in effeciency doesn't outweigh the loss in performance from having more shaders in most situations.

Whatever the reason, and much as I like the elegance of US, I can't see them being a panacea for graphics rendering or they'd be getting more attention than they currently are.

 

[BenQ]

"Not getting it right" can be due to timing. For example, it may be argued that SM3.0+ level pixel and vertex shaders are too distinct still to merge, and thus a drop in per-task efficiency would be unacceptable. If a unified architecture experienced such a drop, it could be attributed as much to timing as to "getting it right" i.e. if one waited for later shader models.

I already addressed the timing issue in my prevous post. IF they "got it right" thn the time IS NOW. And on a side note Xenos has been said to have "Beyond" shader model 3.0 ( see Dave's Xenos article )

Unified shaders may also be right for one system, but not right at this time for another. This harks back to whether "all else is equal".

I don't see how.... IF ATI's unified shader DO function as they claim they do.

As far as the PC space is concerned, it's still even more complex. DirectX does not support unified shading, and when it does, IIRC, it may no longer insist on the shaders being merged on a hardware level. NVidia will have to embrace unified shading from a software perspective, but will they merge on the hardware level, or do so initially? Also, I'm not sure if either company will want to be the first to introduce them in the PC space..there's some speculation that existing and near-future PC games wouldn't work as well with unified shaders as traditional architectures. If that were the case, I think both companies would want to "go together" to avoid risking even a short term drop in (apparent) performance relative to the competition.

Where have you seen speculation that unified shaders wouldn't work well on existing and near future PC games? And what is that speculation based on?

 

[BenQ]

One question that concerns me is if ATi have got it right, why isn't their next PC card using Unified Shaders? Surely if they have that realworld advantage they'd use it?

"Hey, chief! These Unified shaders outperform standard shader by as much as 25% as they're more efficient, and we've got the technology when our rivals haven't."
"That's fantastic. Now let's not use it."

:?

I can think of 2 possibilities.

1. Unified shaders may not work aswell in the world of PC's as they do within the closed box of a console..... but I can't think of any reason why.

2. Contractual obligation. The team who worked on the C1 is not the same team who worked on the R520. M$ may infact own the Xenos and all related technology..... atleast for a time.

I simply can't imagine that Unified shaders DON'T work aswell as ATI has claimed. If they didn't provide better performance reletive to their standard shader designs, then they quite simply would have been SCRAPPED. M$ would realize this aswell and NOT have them in their new Xbox.

 

[Shifty Geezer]

What they lose in performance they gain in flexibility, which probably works better in the eDRAM system as there's points I believe where you want to do just vertex work. Also I'm quite willing to believe ATi are 'flexible' with their figures as any other company. Perhaps that 90% efficiency was only in a very specific situation and the load balancing in real world use is nearer 75% efficient?

All in all I don't expect Xenos to be any more or less powerful than RSX on the whole. I think the two architectures are different, not overall better/worse. What one gains in one area, the other gains in another, and the two roughly balance out. Any system that was developed to get more perforamnce out of the same number of transistors would 1) Have been experimented on by all parties and 2) Be used at every conceivable opportunity. Kinda like Cell versus A64. What it gains in float power it loses in ease of programmability. The limiting factor is number of transistors, and what you do with those transistors determines the strengths and weaknesses of your system.

 

[blakjedi]

M$ may infact own the Xenos and all related technology..... at least for a time.

By george I think you've got it!!!!!:smile:

 

[Titanio]

I already addressed the timing issue in my prevous post. IF they "got it right" thn the time IS NOW. And on a side note Xenos has been said to have "Beyond" shader model 3.0 ( see Dave's Xenos article )

There have been beyond 3.0 GPUs for a while now. The question is, is it beyond enough to make pixel and vertex shading so similar as to make merging them attractive and whether the performance is still there. There's a reason we started at SM1.0 and not SM4.0 Performance has to keep up.

Where have you seen speculation that unified shaders wouldn't work well on existing and near future PC games? And what is that speculation based on?

I had seen discussion of the possibility. It is possible that there could be a transitionary period from one architecture to the next..if "the old way" yielded better performance with current games, it could be temporarily painful for the "pioneer" if they were on their own. Of course, the "pioneer" could hit the ground running and not look back. But there are other reasons why NVidia and ATi are not currently persuing unified shaders in PCs (there has to be if they're not).

I don't see how.... IF ATI's unified shader DO function as they claim they do.

If it does all that is promised, that's great.

The question is: why aren't NVidia using unified shaders now? You may as well ask the same question of ATi however in the PC space. The most obvious answer is: the API does not support it yet.

Now, moving away from PCs, I think your real question is: why is a dedicated architecture better for PS3 than a unified architecture? Or more specifically, why is RSX better for PS3 than Xenos would be? Assuming ATi "got it right", would Xenos necessarily be better for PS3 than the architecture it's getting? Frankly, no, not necessarily. All else is not equal.

I can think of 2 possibilities.

1. Unified shaders may not work aswell in the world of PC's as they do within the closed box of a console..... but I can't think of any reason why.

Neither can I - one of the biggest reasons for unified shaders is to address the issue of utilisation, which would be bigger in PCs than closed boxes.

2. Contractual obligation. The team who worked on the C1 is not the same team who worked on the R520. M$ may infact own the Xenos and all related technology..... atleast for a time.

MS does own the IP for Xenos. Although this presumably would not preclude a different team independently persuing a unified architecture for a PC card, assuming no other clauses preventing that.

I simply can't imagine that Unified shaders DON'T work aswell as ATI has claimed. If they didn't provide better performance reletive to their standard shader designs, then they quite simply would have been SCRAPPED. M$ would realize this aswell and NOT have them in their new Xbox.

Not necessarily. It could be too late by the time you realise shortcomings in an architecture to hit the brakes and go into reverse. If that weren't the case, bad architectures/implementations would never happen, but they do.

 

[Arty]

"Getting it right" isn't just about doing something, it's also about timing. Was putting FP32 precision in NV3x getting it right? Not to draw the comparison - I'm not saying it was a mistake to go with unified shaders now -just that doing something first isn't always a good idea. Some things can be premature.

Premature for now, maybe this year but since Microsoft intends 360's lifespan to be 4-5 years (like every console), I wouldnt call that premature. I'd doubt the traditional approach if it is expected to do the same.

 

[Titanio]

Premature for now, maybe this year but since Microsoft intends 360's lifespan to be 4-5 years (like every console), I wouldnt call that premature. I'd doubt the traditional approach if it is expected to do the same.

To continue the analogy, NV3x isn't any better for FP32 shading (or shading fullstop!) as it was 3 years ago.

To clarify, I'm not saying it's premature to use unified shaders! Just that it can sometimes be premature to come with new features. If ATi have "got it right", great for them.

 

[ShootMyMonkey]

1. Unified shaders may not work aswell in the world of PC's as they do within the closed box of a console..... but I can't think of any reason why.

Neither can I - one of the biggest reasons for unified shaders is to address the issue of utilisation, which would be bigger in PCs than closed boxes.

I think the complexity of things may make a difference. There is surely bound to be some cases where an RSX/G70-type design would outperform a Xenos-type design and vice versa. RSX is clearly skewed towards pixel shading power.

If we're anyway skewed towards pixel processing, there's little to gain by the load balancing of a unified shader architecture (unless the pipelines are every bit as capable in the wide co-issuing the RSX pixel pipes). However, doing this on PC could be a mixed bag because often times, more complex pixel processing means more texture data moved to the card -- and a PC has utter crap bandwidth compared to a console for doing that part of the task. Then again, the PC will, unsurprisingly, often have cards with a lot more memory -- I'm sure there will be 1 GB cards before we know it.

I simply can't imagine that Unified shaders DON'T work aswell as ATI has claimed. If they didn't provide better performance reletive to their standard shader designs, then they quite simply would have been SCRAPPED. M$ would realize this aswell and NOT have them in their new Xbox.

Not necessarily. It could be too late by the time you realise shortcomings in an architecture to hit the brakes and go into reverse. If that weren't the case, bad architectures/implementations would never happen, but they do.

It could just be that this all ties in to the whole WGF 2.0 plans for standardizing unified shader architecture and so on. If Xenos is, to some extent, a proof-of-concept in just how effective the idea can be, chances are that it's far from the best example of the concept, but at least gets the point across.

The necessity of tiling because of the small eDRAM is possibly a potential weakness, but I think in the PC world where nobody seems to balk at the thought of buying pairs of $700 video cards, that wouldn't really be an issue, since they can make a chip with 32 MB of eDRAM and stupidly rich people will pay whatever it costs.

 

[Guden Oden]

I think you just have to look further into nVidi's doublespeak on the issue.

I'm not interested in making this an ATI vs nvidia thing, that is an exercise in futility IMO, and won't lead to any kind of a reasonable debate.

And where has ATI claimed that their unified shaders have "efficientcy issues"?

There was an interview with...umm...well, one of the tech guys of ATI, where he acknowledged that there indeed are performance issues with running vertex and pixel shaders on the same hardware just like nvidia said. This isn't exactly a secret you know, pixels and vertices have different needs and different demands, which is a large reason why they run on different designed hardware in the first place!

Anyway, sorry I can't provide you a link or anything to that (rather brief, as I recall) interview , I'm not google or the wayback machine, I don't keep the entire intarweb stored in my brain.

Everything I have read suggests that the WHOLE POINT in moving towards Unified shaders is to increase efficientcy beyond what standard shaders are capable of, due to their very architecture, and what ATI is claiming is that they are damn near 100% efficient.

I think maybe you've got caught up in marketing speak. I don't see how a piece of hardware would suddenly become "damn near 100% efficient" simply by running all shading on the same pieces of hardware. Unless of course you by efficiency mean you would have less stalls due to either type of shader hardware getting bogged down by too much work, leaving the other type idle. That situation would of course not appear in a unified architecture since there is only one type of shader processor implemented, so there's nothing else to wait for.

I don't think a first-gen unified shader however is any more likely to achieve real-world performance that is near or at theoretical maximum than a third or fourth generation traditional architecture however, and likely it would actually be a little less likely at hitting max performance actually due to being a first implementation.

but I can't accept that a ballpark figure for real world specs is impossible to know. Even a ROUGH ballpark figure would HAVE to be more accurate than theoretical maximums.

But a rough ballpark figure of WHAT?

Look at farcry for example, even within a single title, you have levels that are either CPU limited or GPU limited, or levels that in part are CPU limited or GPU limited in different parts of the level. And this picture changes depending on which features you enable and which patch of the game you run, as they introduce new rendering tech in later versions.

 

[coldstorm]

But with the slew of rumour on the net that G80 is indeed a unified pixel architeture becuase of WGF 2.0. Opengl 2.0 which is proberly ps3 main 3d api. it would make sense to go with the common architecture of pixel shader and vertex shaders and not go and try something new like ati did when making the xenos for 360 as head start in wgf 2.0 compliance.