WTH IGN?! They posted the spec analysis from Major Nelson!!

[Acert93]

http://xbox360.ign.com/articles/617/617951p1.html

That is just so wrong in so many ways. I mean, letting PR folks confuse the average gamer even more

I mean, they just pasted the entire thing It is good that the general processing power of 3 PPCs is getting noted (I think we get blinders at times), but that entire thing misconstrues the entire arguement! For every nugget there are 10 things wrong. They screw up the CELL info, they do an apples-to-elephants bandwidth comparison

Man, PR has ruined this E3 for me from both Sony and MS. All I wanted was some nice PLAYABLE next gen games and no one showed up.

At least Nintendo did not get involved... heck, if they provide a killer system in the same league I may just get that to spite MS and Sony

 

[Tacitblue]

One word, FUD. Even with the controversy over what was prerendered and what was realtime, I got the impression that the best things i saw were real time, and i'm trying to be objective, the Nvidia Luna demo, the Unreal 3 one as well, those looked qualitatively better than what was playable on the floor. (Yes I know 360 also has an U3 engine game), 6 months behind Microsoft, nothing playable per se, but capable of demonstrating that it works, and works well. Who'd not want to get the damage control out? To be fair, there were a lot of interesting games that MS is working on, which when they reach maturity are definately going to be under Christmas trees.

 

[jvd]

posted by you in another thread no need to have both

Quote:
Another bit of information sent our way is the final transistor count for Xbox 360's graphics subset. The GPU totals 332 million transistors, which is spit between the two separate dies that make up the part. The parent die is the "main" piece of the GPU, handling the large bulk of the graphics rendering, and is comprised of 232 million transistors. The daughter die contains the system's 10MB of embedded DRAM and its logic chip, which is capable of some additional 3D math. The daughter die totals an even 100 million transistors, bringing the total transistor count for the GPU to 232 million.

http://xbox360.ign.com/articles/617/617951p3.html

And the image scaler seems to be a 3rd chip according to the Anand / Firingsquad interviews.

This makes for an interesting question of tradeoffs. Since the eDRAM should be fairly small per-transister, I am guessing the chips should be about the same size. Some questions (no real answers yet).

1. Since the R500 is 2 dies, will this have a positive effect on product yields? (It was noted that the R500 has a lot of backup redundancy so if some of the cache or ALUs are bad in production units there are some extra just in case to help yields, chips that have all of them working will have the extras disabled)

2. Since they are about the same size, was the Unified Shaders/eDRAM a good trade off compared to the more traditional GPU like the RSX?

 

[Acert93]

Well they ARE very different pieces of info; IGN just tacked that on at the end of the Major Nelson thing (it did not even come from him).

I thought the questions I asked also were quite different. This is sure to be a MS flame fest and locked

The Transistor one answers a lot of questions.

Oh well.

 

[jvd]

keep them together . Its easier and since its the same article alot will be repeated twice (or three times since this was already talked about )

 

[Acert93]

Not a problem. That is why I posted it separate though... the transistors is new... the Major Nelson is not (just the IGN is posting it! egad!).

I was afraid the Major Nelson stuff would scare away the people who would actually want to talk intelligently about the transistor information

 

[DemoCoder]

Talk about bullshit. He claims that it is CELL+RSX which is capable of 51 billion dot products/s, when in fact, it is RSX alone, and this is clear from Nvidia's presentation.

And I love the "278Gb/s vs 22Gb/s", which is wrong on two counts. First, because the RSX can access both GDDR and XDR memory. And second, because of the phony bandwidth numbers, comparing effective with real instead of effective vs effective, and not looking at the situation holistically, since there is more to bandwidth than AA-framebuffer bandwidth.

 

[Acert93]

And I love the "278Gb/s vs 22Gb/s", which is wrong on two counts. First, because the RSX can access both GDDR and XDR memory. And second, because of the phony bandwidth numbers, comparing effective with real instead of effective vs effective, and not looking at the situation holistically, since there is more to bandwidth than AA-framebuffer bandwidth.

I agree that it must be looked at holistically and it is bunk.

But I did think Dave had mentioned that the bandwidth between the eDRAM and the logic on the eDRAM was a *real* 256GB but he was going to check on it. The BW between the eDRAM die and the Shader die is only 256GB/s effective (32 write, 16 read I believe).

 

[aaaaa00]

Talk about bullshit. He claims that it is CELL+RSX which is capable of 51 billion dot products/s, when in fact, it is RSX alone, and this is clear from Nvidia's presentation.

Do you have a better link than I do?

I thought it was far from clear -- the slide I saw from the Sony presentation looked like this:

Now, why would the slide be titled "Most Powerful Graphics SYSTEM ever built" and not "Most Powerful GPU ever built"?

Why is there a picture with BOTH CELL and RSX on the slide?

Why does it say "512 MB of Graphics Render Memory" instead of "256MB Graphics Memory"?

Why does the slide say "2 TFLOPs" when from other sources, Sony's been claiming 2 TFLOPs is system-wide?

Unless you have better evidence, I find it hard to believe Sony would pass up adding everything the CELL can do into the GPU numbers in order to make itself look better.

 

[jvd]

Doesn't the rsx have to acess the xdr through the cell chip ? which will greatly impact the transfer rate and the use of it

 

[Acert93]

Doesn't the rsx have to acess the xdr through the cell chip ? which will greatly impact the transfer rate and the use of it

It has access through tThe FlexIO on CELL which is really fast and designed to shift information around. It is insanely fast actually, with the GPU having like 20GB/s write and 15GB/s read to the XDR (or something like that).

 

[aaaaa00]

GPU in x360 has direct access to the CPUs L2 cache at like 20+ GB/s as well, so things like vertex pushes can come from it without making a round trip to main memory, if the developer so wishes.

You could have one of the cores generating vertices on the fly and push it out from the CPU without eating main memory bandwidth, for example.

 

[Agisthos]

MS would not have released this if Sony hadn't tried to hype the concept renders as ingame and say the PS3 is twice as powerful.
Sony started it and MS is now obviously engaged in a 'gloves of' dirty PR battle.

So its not all one way traffic guys.

The document is obviously biased and wrong in many areas but it should not be dismissed completely - the CPU architecture comparisons bare some discussion at least.

The GPU comparison is just wrong because of one simple thing, it presumes the RSX is just a twice as fast 6800 ultra. Nobody knows what cool features will be in this chip it will be totally different to the 6x series.

 

[Acert93]

GPU in x360 has direct access to the CPUs L2 cache at like 20+ GB/s as well, so things like vertex pushes can come from it without making a round trip to main memory, if the developer so wishes.

You could have one of the cores generating vertices on the fly and push it out from the CPU without eating main memory bandwidth, for example.

Ewwww good point. I remember reading this in a patent.

Really, someone should start a thread to collect basic system information + design quarks like this. Not a compare/constrast necessarily, but something to highlight the system design, strengths and weaknesses, and just be a nice reference.

Actually I think I may do that

 

[Panajev2001a]

Talk about bullshit. He claims that it is CELL+RSX which is capable of 51 billion dot products/s, when in fact, it is RSX alone, and this is clear from Nvidia's presentation.

Do you have a better link than I do?

I thought it was far from clear -- the slide I saw from the Sony presentation looked like this:

Now, why would the slide be titled "Most Powerful Graphics SYSTEM ever built" and not "Most Powerful GPU ever built"?

Why is there a picture with BOTH CELL and RSX on the slide?

Why does it say "512 MB of Graphics Render Memory" instead of "256MB Graphics Memory"?

Why does the slide say "2 TFLOPs" when from other sources, Sony's been claiming 2 TFLOPs is system-wide?

Unless you have better evidence, I find it hard to believe Sony would pass up adding everything the CELL can do into the GPU numbers in order to make itself look better.

Even if we add the Dot4 operations from the Broadband Engine, I think we are selling RSX short if we compare things like that document did.

If they add the external EDRAM block's transistor's count passing it all as logic, I can do the same for RSX: it is now using 2.35 Billion Transistors .

 

[Acert93]

If they add the external EDRAM block's transistor's count passing it all as logic, I can do the same for RSX: it is now using 2.35 Billion Transistors .

Panajev2001a you know better than feeding the trolls!!

 

[DemoCoder]

Yeah, you're right, I was thinking of the fact that it can do 52-dots per clock.

Overall tho, he's still selling bullshit. It would be correct to say that the XBGPU's 48Gb/s bandwidth is enough to achieve its peak fillrate of 4Gpix/s, whereas the RSX probably will be stuck around half that unless it can render separately to XDR and GDDR (e.g. Z-buffer in GDDR, FB in XDR)

But this is irrelevent, since the peak color+Z fillrate will never be reached on either system, since color rendering will involve shaders, and we would expect most rendering which is writing shaded pixels to be very shader bound.

Instead, it is the peak stencil/z fillrate which matters, and there, the systems are probably about equal.

 

[blakjedi]

Talk about bullshit. He claims that it is CELL+RSX which is capable of 51 billion dot products/s, when in fact, it is RSX alone, and this is clear from Nvidia's presentation.

Do you have a better link than I do?

I thought it was far from clear -- the slide I saw from the Sony presentation looked like this:

Now, why would the slide be titled "Most Powerful Graphics SYSTEM ever built" and not "Most Powerful GPU ever built"?

Why is there a picture with BOTH CELL and RSX on the slide?

Why does it say "512 MB of Graphics Render Memory" instead of "256MB Graphics Memory"?

Why does the slide say "2 TFLOPs" when from other sources, Sony's been claiming 2 TFLOPs is system-wide?

Unless you have better evidence, I find it hard to believe Sony would pass up adding everything the CELL can do into the GPU numbers in order to make itself look better.

Spot on aaaaa0.

 

[blakjedi]

Yeah, you're right, I was thinking of the fact that it can do 52-dots per clock.

Overall tho, he's still selling bullshit. It would be correct to say that the XBGPU's 48Gb/s bandwidth is enough to achieve its peak fillrate of 4Gpix/s, whereas the RSX probably will be stuck around half that unless it can render separately to XDR and GDDR (e.g. Z-buffer in GDDR, FB in XDR)

But this is irrelevent, since the peak color+Z fillrate will never be reached on either system, since color rendering will involve shaders, and we would expect most rendering which is writing shaded pixels to be very shader bound.

Instead, it is the peak stencil/z fillrate which matters, and there, the systems are probably about equal.

isnt thr r500, due on the efficiency of the of the edram, supposed to sustain near peak throughput and utilisation?

 

[DemoCoder]

It won't when HDR rendering is used, which if you listen to the hype, will be used in almost every game, certainly in Unreal Engine games.

But peak pixel throughput is really kind of irrelevent. If the future is HDR, shader bound, and global shadowing/lighting, that means the deferred rendering scenario is the most likely, which means you're effective zixel fillrate will be important as well as your HDR performance.

Of course, the R500 also won't reach its peak fillrate if it has to swap out the framebuffer, since there is no leeway in that 48Gb/s bandwidth, and copying framebuffers around puts them over the limit, meaning some ROPs will be waiting on memory.