GPUs vs Cell benchmarked

[Mintmaster]

Not to sound too dumb, but how could this new information help PS3 devs make PS3 games better?

If you read up on fractals, you'll know that there is little to no connection between this example any real-world situation I can think of.

Fractals like the Mandelbrot and Juliet sets repeat a few calculations on a few variables for many iterations. You basically need no data going in and out of CELL for the bulk of the algorithm, so all the things that can degrade CELL's performance - in order execution, memory access predictability - don't apply here.

Even branching, which is normally tough to predict for raytracing as you descend hierarchical spatial structures, is a piece of cake here. Just assume the loop will keep looping.

This is basically CELL in all it's glory, much like the 200GFlops matrix multiplication.

 

[flick556]

If you read up on fractals, you'll know that there is little to no connection between this example any real-world situation I can think of.

Fractals like the Mandelbrot and Juliet sets repeat a few calculations on a few variables for many iterations. You basically need no data going in and out of CELL for the bulk of the algorithm, so all the things that can degrade CELL's performance - in order execution, memory access predictability - don't apply here.

Even branching, which is normally tough to predict for raytracing as you descend hierarchical spatial structures, is a piece of cake here. Just assume the loop will keep looping.

This is basically CELL in all it's glory, much like the 200GFlops matrix multiplication.

I don't doubt that these types of programs are ideal for cell. Yet I also think that these types of programs can be useful in games or other applications.

 

[ihamoitc2005]

Geometry

That's what i thought. Cell is a geometry monster, which we all knew already. Problem is that RSX will almost NEVER be geometry limited as it will be fillrate or bandwidth hungry much sooner than that.

So Cell will have to be used for something other than geometry if it is to really help RSX.

With closed box setting developer has freedom to maximize geometry use of PS3.

 

[Fafalada]

Problem is that RSX will almost NEVER be geometry limited as it will be fillrate or bandwidth hungry much sooner than that.

Ok I'll bite - how did you arrive to this conclusion?
Generally it's the software that determines what the utilization is, and last I checked we haven't seen any real performance characteristics for PS3 software yet. Heck, one could argue we haven't seen any real PS3 software yet.

 

[MfA]

If you read up on fractals, you'll know that there is little to no connection between this example any real-world situation I can think of.

That could have been said whatever GPGPU application they had benchmarked against Cell, even raytracing, and the outcome would almost certainly have been the same. GPGPU research is a little like most robot research, a good way to stretch your engineering mussles ... but best not look too closely at the scientific and practical value (Not going to make friends with this.)

 

[Jawed]

If GPGPU video transcoding turns out to be the real deal then it'll be a little more than stretching engineering muscles.

That would make for an interesting comparison of Cell and GPU...

Jawed

 

[nAo]

(Not going to make friends with this.)

Don't worry, I think they know that anyway so yes, I strongly agree with you.

 

[Mintmaster]

That could have been said whatever GPGPU application they had benchmarked against Cell, even raytracing, and the outcome would almost certainly have been the same. GPGPU research is a little like most robot research, a good way to stretch your engineering mussles ... but best not look too closely at the scientific and practical value (Not going to make friends with this.)

True, and I'm just answering his questions. Still, GPGPU could at some practical applications, just not for games with the exception of simpler examples which aren't usually classified as GPGPU.

Regarding the comparison, at least they could have picked an example with some data flow.

The way this example did ray tracing was similar to parallax mapping with distance functions, except the distance functions were analytical instead of stored in a 3D texture, meaning many more cycles for the 7800 and many, many fewer for CELL.

 

[MfA]

If GPGPU video transcoding turns out to be the real deal then it'll be a little more than stretching engineering muscles.

Maybe for sloppy transcoding which just directly converts video using the already encoded motion ... but that is trivial anyway. As for encoding, you will only ever get second rate quality ... the R/D optimization in good encoders tie every major part of the encoding process together very intimately in a way not suited to a GPU. Even offloading ME to the GPU doesn't make much sense, since the GPU is only suited to antiquated ME algorithms ... at best you could use it to get some motion candidates for the real ME.