Xbox 360 running Ruby at Siggraph, Pictures

according to dave's Xenos article, it may be until next year until a desktop part reaches the Xenos real time performance.

Will it surpass it? probably, unless, we are underestimating the EFFICIENCY of the Unified shaders on Xenos.

 

I did nothing of the sort, that was hasanahmad..

The INQ article doesn't say that either

 

The Xenos almost certainly won't be able to compete with the R600.. We are talking about a card released 1 year from now. And if you want to believe ATI GPU's won't be as efficient next year as it is now be my guest

 

R520 was scheduled to release Spring/Summer 2005 (before it was delayed by about 3 months) and Xenos is scheduled for Winter 2005. Naturally Xenos has a longer lead-time from production-ready to street-date because it's a console.

The net effect is that both are effectively concurrent releases (within 3 months of each other). R520 is no way a "6-months more-advanced" part.

Also just comparing architectural generations, Xenos is about two (arguably three) whole generations ahead of R420 meaning there's plenty of space in the gap (between R420 and R600) for R520 and R580.

R600 will indeed be more advanced in shader tech than Xenos, but Xenos will be far in advance of R520 and RSX. The jump from SM2 to SM3 (R420 to RSX/R520) is quite minor compared to the jump from SM3 to SM4 (R600). Xenos is beyond SM3, but we don't know yet how close to SM4.

Jawed

 

All the signs point to R600 having the same conceptual design of shader core as Xenos (i.e. unified for both vertices and fragments) - but with increased functionality.

The one thing we can be sure R600 won't have is EDRAM for the framebuffer - ATI has said so.

Jawed

 

Damn that is one sexy console. Did they ditch the silver plated disc tray. i hope so, it looks better just plain white like the console.

 

True enough, and I expect R520 to be on the market before X360. But manufacturing for likely Xenos started earlier, and it certainly taped-out earlier. But yeah, there was a delay for R520. Either way, I would expect it's raw shader/programmable power to be in the same ballpark, if not exceed it. Utilisation is another issue, it obviously hasn't got eDram either, but I would expect more complex chip coming a little later to hold some advantages.

The advancement has more to do with utilisation/flexibility than capability/power. The unified ALUs in Xenos are SM3.0+, from that perspective I would not consider them, for the individual tasks of pixel and vertex shading, to be "far in advance" of R520/RSX.

 

You still don't get it do you? - R520 is less tech and was designed to be released/manufactured earlier than Xenos.

I'm fully expecting Xenos to be way faster than R520 and I'm expecting R520 to show G70 a thing or two, too.

MEMEXPORT, all on its own, puts Xenos in a different league. Xenos can tessellate - RSX (G70) cannot.

Jawed

 

The r520 is based on the r300. Which already is over 3 years old . The xenos is based on the r400 which was abandoned thus its design is newer .

The advantages I see for the r520 is mostlikely more pipelines 24 of them most likely . Which will create more fillrate .

However both designs will be based on sm3.0 .

The xenos advantage comes from being much more efficent , edram and unified shaders

The xenos will also most likely never be called to render above 720p with 4x fsaa which is what it was designed for . The r520 will most likely be asked to run 1600x1200 with fsaa 6 or greater .

I agree with jawed that hte r520 will be slower than the xenos . However this is only when asked to do what the xenos was built to do . I suspect that for the aplications on the pc the r520 and the g70 will be both better suited

 

yes, R520 is still based somewhat on R300 tech, while Xenos is basically a completely clean sheet design, AFAIK


also, time tables are not necessarily indicative of performance .. AFAIK, the main indicators of performance, are: 1. design , and 2. transistor budget ......... i.e.: how many transistors you have, AND how you use them ............................... time is only a factor because depending on the time period, different fabrication processes may be feasible to use (i.e.: 90nm vs 65nm), which effects what your transistor budget can be, because of cost considerations, etc.

also, I recall this statement:

[source: http://www.extremetech.com/article2/0,3973,1220430,00.asp ]

now, as it turns out, the performance of Xenos has a chance to be higher than that of Revolution's GPU, even though Revolution's GPU seems to be the 'later' product ......






However, I do agree with you, Titanio, that the raw shader power on the R520 spec. sheet may be higher that that on Xeno's spec. sheet ... but, it remains to be seen whether the real world performance may be a different story........

 

People should stop confusing high efficiency with high performance. If R500 is 100% efficient compared to R520's 80% efficiency, but R520 has 50% more theoretical power, then R520 will have the greater effective power.

I mean, who is actually going claim that the more generalised combined shaders are going to be at least as powerful as more specialised dedicated shaders? Certainly nvidia says the exact opposite. Are they lying? Well if they are then why arn't they using unified shaders already? For that matter why is ATI only using unified shaders in the one of their two platforms that can't be benchmarked?

As for the talk about R400... well it was cancelled in favour of R420 because it was too slow. R500 is R400's successor, R520 is R420's successor. Do the math.

 

OK, I'll put it simply - if you "averaged out" Xenos to be a 16VS+32PS design, it's 100% faster than R420 - and that's before you take the fact that it can run shader code with approximately 50% greater utilisation per ALU (60% averaged utilisation for R420, 95% for Xenos).

So Xenos is in the ballpark of 150% of the performance of R420. And that's purely in shader code.

Currently I expect R520 to be around 30-75% faster than R420...

Jawed

 

That may be true . However what if xenos has 95% efficency , r520 has 60% efficiency and has only 20% more raw power ?

Not tomention that there is just more than fillrate and polygon performance.

Nvidia is clearly heading to unified shaders . I don't know why you don't believe this . Nvidia has said it already . THe only reason why they aren't talking it up right now is because they don't have it . In the pc sector right now there is no reason to have it , but with dx 10 there will be a reason and you will see nvidia support it .

Sorry can you post a link to why it was canceled ? You claim because it was slow. Yet that is not what was said or hinted at by many insiders . So unless you can back that up you should remove that comment .

The r400 wasn't made because it wasn't feasible on 130nm and 110nm . 130nm being what was avalible at the time of the r420 launch . However on 90nm they seem to have been able to make it just fine

 

We've already discussed that the 1.5 million polygons displayed in realtime is a bogus. They were talking about the source model for the normal maps, the actual models are quite low-resolution.

 

Umm... why? R420 has 16 pixel shaders each with two Vec4 ALU's and then another 6 vertex shaders with 1 Vec5 ALU each (I know they arn't single ALU's but for the sake of simplicity...) So that gives us (for the sake of simplicity) 38 Vec 4 ALU's vs 48 Vec5 ALU's in the R500. So its not even close to 100% faster.

Where do those efficiency figures come from? As I understand it, the USA is more efficient because all the "pipelines" are always active. In a normal architecture which is pixel shader rather than vertex shader limited, you would have.. what... 33% of your vertex shaders idle? In the R420 that 36 ALU's active and 2 inactive (from the vertex shaders), thats a lot greater than 60% efficiency.

Which would be far higher than any graphical leap in the same time period ever. Given that a console has power, heat, and cost restrictions compared to the PC, and ATI have been "apparnetly" unable to match that performance in another one of their parts without those restrictions, wouldn't you agree that these performance predictions are a little optimistic?

 

The latter may be true, but there should there not be more silicon in R520's shaders than Xenos's?

This seems somewhat contradictory..If R520 was designed for a spring/summer release, shouldn't something designed for late summer have an advantage, all else being equal? (I'm assuming here the G70 was on time, I'm not aware of its history, but on the face of it, if that were the case, that's a little contradictory..)

Lovely numbers you've got there

You could "average out" Xenos to a 16VS+32PS design? I don't think so. Those ALUs do not a (R520/G70) pipeline make.

If you were to draw equivalencies based on floating point power, Xenos wouldn't even be the equivalent of 8+24, based on the G70 design (@500Mhz). Although their floating point power is arranged differently, and such comparisons are apples-to-oranges of course..any kind of "normalisation" is not going to be 100% accurate, and probably should be avoided.

 

Only ONE Vec4 ALU - the other ALU is some kind of mini-ALU - there is no dual-issue in R420 like there is in NV40/G70 (well that's the current theory - per-clock performance comparisons of R420 and NV40 tend to agree, as far as I can tell).

I'm quite happy calling them Vec5 ALUs, for the sake of this comparison - as long as we remember the scalar part of both VS and PS ALUs provide the option for a co-issued scalar instruction.

Teehee, well as you can see I'm working on the basis of R420 being 6VS+16PS = 22. R420 is effectively one ALU per pipe with the ALU being capable of co-issues. There are some extra co-issue capabilities in there - e.g. it's documented that R420 can co-issue 5 scalar operations in a VS ALU.

ATI says that current GPUs are 50-70% efficient - I plumped for 60% for the sake of simplicity. 95% is Xenos's efficiency according to ATI.

No, you're missing the fact that Xenos never waits for dependent instructions, whether that's dependent texture operations or simply serially-dependent ALU instructions. There's also no waiting for surface texturing (e.g. with anisotropic filtering) of pixels, regardless of batch size - even 1 pixel triangles can be textured without a stall.

The concensus is that Xenos is 2 generations ahead, not one. Less than 150% over two generations would be embarrassing, to say the least.

Anyway, Xenos doesn't have a "base". You can't say its starting point is R420.

No. The power/heat/cost thing is taken care of by Xenos being fairly slow on 90nm, and not being a huge device. Bear in mind that back when M$/ATI started XB360, they were prolly expecting it to launch on 65nm tech. 90nm is way "late" by guesstimates from way back when. Also, we don't know what proportion of Xenos is redundant (for yield purposes) - which also affects heat. Is the quoted transistor count for Xenos including the redundant parts?...

R600 is going to be a 90nm part too. It is prolly 6-9 months behind Xenos (Vista was going to be mid-2006, not end, for quite a while). It'll have to be much bigger than Xenos (parent die) due to full SM4 functionality and prolly requiring backward compatibility (DX8 fixed-function hardware, etc.).

Jawed

 

OK, you've got me there - why on earth is the quantity of silicon relevant?...

In what possible way could "all else be equal"? Seriously. Xenos is a world apart from any preceding PC GPU. About the only commonality it has with previous GPUs is the fact it's an immediate mode renderer - and even then it has some deferred rendering tricks up its sleeves

I'm being very careful here to make a comparison of Xenos with R420 in terms of pipeline capability. We can then extrapolate from R420 to G70 via NV40 if you like...

Sigh - but they do make R420 pipes - in rough terms of instructions per clock. See my previous reply about R420's lack of dual-issue... Xenos doesn't dual-issue either, by the way.

(Actually, R420 and Xenos can dual-issue - but the instructions are split into texturing and non-texturing in the dual-issue.)

Which is why I used R420 - R420 is effectively a one-ALU per pipe design. The reason NV40's dual-issue capability doesn't totally floor R420 is that NV40 can hardly ever dual-issue. G70 fixes that problem, somewhat - but there are myriad scenarios when it cannot dual-issue either.

Yes, there are all sorts of near-SM4 performance boosts in Xenos which will take it way beyond the 150% boost over R420

Jawed

 

This is the same Ruby demo ATI showed at E3 on the only Beta hardware they had on the show floor at the time (all X360 demos ran on Alpha at the show, save for the Ruby demo).

If I was at home, I would upload and show you the pics from it.

Also, this ruby demo was one designed for R520, but got hijacked and recoded in 2 week period to show off the X360. The demo ran @ 30fps, and without any AA at E3 though. I wonder if ATI tweeked it further for Siggraph....