Did you realize that the first thread you quote is from 2005.? The PS3 wasn't even out yet hell the xbox 360 wasn't even out yet,many of those claims have been proven wrong already by the way. The second one,basically confirms that Cell is a beats that can do pretty much anything,every single thing he points there is done on 360 by the GPU,in other words taking GPU cycles,and lowering maximum fill rate,there are several things on that list like Branching which were dismiss long ago. The fill rate of the 360 is almost double of the PS3 on paper how do you explain games like Uncharted 3 topping 360 games.? Either those specs are a lie or something doesn't allow the 360 to get even close to those 500 million polygons,is not magic the xbox was the same 125 million polygons but in real life with effect an all it was more like 30 million. A developer claimed that there was no way to get interaction between the different SPUs of the Cell. This heralded Sony to come in and teach a miniature lesson about what does and doesn't work with the Cell's SPUs. The interesting tidbit spoken about was branching AI on the Cell. We'll use a quote here so we don't mess up the interpretation: "Branching is a common technique...where a program randomly chooses a few samples from a larger set of options, and then tests each to see which is the best...Most developers have claimed that the SPUs would be absolutely terrible for branching. As Sony put it however, branching is absolutely terrible for ALL processors. In their experience, they said, it is less terrible for the SPUs however. In the upcoming game Heavenly Sword, they said that moving the branching AI off of the Power Processor Unit (PPU) increased the performance of that particular process. In other words, the same branching ran better on the SPUs." http://www.joystiq.com/2007/03/09/sony-says-cell-can-handle-branching-for-ai-better-than-the-re/ If Branching is bad for Cell imagine how bad it would be for the xbox 360 CPU.
tormentos
To support Xenos's high polygons count, look up hardware tesseleation. Most multi-platform games doesn't use Xenos' hardware tesseleation e.g. PC development tool/content chain got hardware tesseleation support with DX11.
On "Uncharted 3 topping 360 games" subject, your comparsion statement is subjective. You don't benchmark with disimilar programs .This is like comparing 3DMarks Vantage scores on AMD Radeon HD against Metro 2033 scores on NVIDIA Geforce CUDA.
PS; one should not mix Bechalor Art (mostly subjective) students with Bachelor Science students.
On code branching issue, PPE or SPE's branching unit it less capable than "fat" Out-Of-Order CPU e.g. Intel Core 2.
PPE's branch unit about the same level as Intel Atom. PPE in CELL is nowhere near PowerPC 970FX. This is like comparing Intel Core 2 against Intel Atom. IBM PPE CPU core is about same same level as Intel Atom/ARM Cortex A8 i.e. two instruction issue per cycle with in-order processing processors.
My tablet's AMD bobcatCPU i.e. has two instruction issue per cycle with out-of-order processing. ARM Cortex A9 CPU has two instruction issue per cycle with out-of-order processing e.g. Sony Vita, Apple iPad 2.
PS; Sony Vita's has similar GPU design to AMD Xenos i.e. unifed shader, decoupled texture and shader units design.
----------------------
AMD Xenos has it's own design limitations, but let's address the NVIDIA RSX design.
http://www.eurogamer.net/articles/digitalfoundry-metro2033-article
"For me personally, the PS3 GPU (they like to call it RSX for some reason) was the safe choice because I was involved in the early design stages of NV40 and it's like a homeland: RSX is a direct derivative of that architecture. Reading Sony's docs it was like, 'Ha! They don't understand where those cycles are lost! They coded sub-optimal code-path in GCM for that thing!' All of that kind of stuff..."
G70's code name was NV47. RSX's architecture already existed with NV40.
Are you claiming NVIDIA G7X/RSX has decoupled texture unit and shader/math unit design before G80? Show me the docs to prove NVIDIA RSX has G80's key design improvments.
On Jawed's post
Dependent texture fetches in Xenos (I presume that's what the third point means), will work without interrupting shader code - again RSX simply can't do this, dependent texturing blocks one ALU per pipe -
Jawed's statement refers to G7X/RSX's couple texture and shader/math design.
The NVIDIA G7X/RSX pixel shader and texture unitblock.
![](http://techreport.com/r.x/geforce-7800gtx/pixel-block-small.gif)
ALU pipeline is coupled with texture unit. Texture fetch will stall "FP32 Unit 1", which in turn stalls the rest of the ALU pipeline.
The first NVIDIA decoupled texture unit and shader/math unit design was G80.
![](http://i328.photobucket.com/albums/l327/encia/Decoupled.jpg)
This is where NVIDIA's G80 Giga-Thread technology comes in.
Mordern GPU designs follows the flat ALU (stream processor)design.
![](http://media.bestofmicro.com/Radeon-HD-4850,1-E-112802-13.jpg)
AMD Radeon HD's stream processor block. NVIDIA G80 and AMD Xenos follows this topology.
Flat stream processor design for NVIDIA G80.
![](http://i328.photobucket.com/albums/l327/encia/USA.jpg)
Log in to comment