A GPU is a highly parallel architecture where they abandon decent generic performance in order to gain performance on a specific set of algorithms related to the raster scan-line algorithm ... Depending on what you were trying to achieve your performance can be far higher by using a second CPU (and an integrated GPU) than by including a dedicated Graphics Card.

When it comes to systems like the XBox 360 and PS3 which were designed to solve (basically) the exact same problem, the architectural differences between the systems are not important.

Except the Cell isnt parellel enough to really shine against GPUs in their specialised fields. Its probably more efficient to have a bigger GPU than to use the Cell to do some GPU work. The original Cell was a 1:6 design which was small enough to be more practical to manufacture, and it all likelihood it would have given them extra budget to make the GPU more powerful anyway.

If we're talking about coming close, then we have Gears 2 - and anyone who things that KZ2 is lightyeras beyond it is just a fanboy. And guess what gears has? Split Screen, something KZ2 couldn't handle. You know what else Gears 2 had? A much smaller development budget. KZ2 was given limitless funds to be tech demo, show what the PS3 can do. Unfortunately all it's proving is that: With a MASSIVE budget, and a VERY long development time, you can make a graphical beast that is missing key features of the genre.

How much better do you think Gears/Halo would look if they didn't have to code in Split screen/Coop?

Not much, since Halo 3 isn't even 720p due do the GPU struggling with HDR. Now if you said 'how much Halo would look better if they cut back HDR/draw distance' it would make more sense.

And Gears is already maxed out. They've cut sharpness and distance draw to bring more detailed textures/characters on Gears 2 which means backgrounds on Gears 1 look way better.

On the other hand Killzone 2 display on native 1080i, which needs 20% more bitrate than 720p on the same quality (link)

Not all stages of a game pipeline are naturally serial in nature, nor are the portions which are easily made into parallel operations cheap, or easy (or even realistically possible) to put on the 360 or PS3 GPU.

GPUs don't tend to handle the animation portion of a game frame, and that's not only excessively expensive (and it scales with the number of characters and skeletal bones you have in your animation, as well as the number of animations you are blending at once), but also something the PS3's SPUs excel at -- as in they can individually process this information (often hundreds of quaternion transforms interacting for each character, each frame) quite a bit faster than a single core of the Xenon can, and on top of that, there are 6 of them available for use, and they can do this while the PPU is off doing something else.

Animation and physics are probably the two largest expenses (by a longshot) of a typical shooter game's CPU frame. They have to be completed each frame before the skinning can happen on skinned meshes (basically all characters), and then sent off to be rendered by the GPU. There's absolutely no question that the Cell completely demolishes the Xenon at these kinds of operations, when they are written in a SPU-friendly manner. Less time preparing info for the GPU == more time for the GPU to render the frame. Yet another substantial rendering advantage of the PS3, and a game engine authored to use the PS3, that has to do with the system architecture, and not the GPU at all.

lol..

umh..yeah but not split screen, and no coop..

Depending on whether or not a workload lends itself to data level parallelism, task levelparallelism, instruction level parallelism, or is just inherently sequential you will potentially get order of magnitude differences in performance with different architectures which have the same transistor count. It may be that the best possible performance (lets pretend this is well defined) of any concievable game will be similar between the xbox360 and the PS3. However, I don't consider a metric with this much room for error to be very useful.

Much of this debate seems to ignore the possibility that the rival platforms will each have significant advantages in different types of games.

The potential advantages of the PS3's architecture go beyond the exploitation of task parallelism. Depending of the FP to MemRef ratio an SPE could potentially have a significant advantage due to the proximity of the local store to the ALUs and FPUs. Of course this advantage could potentially be negated depending on the spatial and temporal locality of those references.

I am on an NSF compiler grant for doing cell research at the moment and although my research is more in the HPC realm, my intuition is that it would be very unlikely that no games would be able to make use of it's advantages with certain applications.

Off the top of my head I would be surprised if a game like total war could not take advantage of the cell allowing significantly larger numbers of troops on the battlefield simultaneously.

There are a number of workloads which do not lend themselves to data parallelism which can take advantage of task parallelism. Although, it is true that if you have the choice between the two, data parallel architectures do tend to have significantly higher throughput.