cell v gpu & gpgpu hypothetical question

[ebola]

How many spu's would it have taken to match a gpu.

I gather sony went through various permutations of parts of rendering done on the cell before settling for a conventional gpu..

how efficient/in-efficient would it be, in terms of rendering performance per transistor... versus a gpu. (how much goes on the shader units - which i presume gather the Cell could probably match - versus say the specialized paths in the texture-read units and all the rasterization side.. )

I was just thinking about that what with talk of an HLSL compiler for ps3's.

Although it makes no sense in the current GPU era, the concept of software rendering on cell/(or similar) fascinates me...

 

[Shifty Geezer]

It has been discussed a lot. Go searching and you'll find some long discussions I think. In summary, in vertex work SPUs are pretty good, but in the pixel rendering things won't be straightforward. We mused an alternative approach to texturing that should be more SPU friendly. Until someone tries to make an optimized Cell renderer, we won't really know. Well, we'll know that it won't really compete! No matter how optimized the software side, GPUs are good at what they do! But Cell should be fast enough for things like a handheld, with a single processor managing code and graphics.

 

[Arun]

But Cell should be fast enough for things like a handheld, with a single processor managing code and graphics.

It would still have lower perf/mm² and lower perf/watt, though, but yeah - the results would be less disastrous than for a non-handheld part.

 

[archangelmorph]

It would still have lower perf/mm² and lower perf/watt, though, but yeah - the results would be less disastrous than for a non-handheld part.

Than a GPU yes but a handheld could never really utilise a full-sized GPU + CPU anyways due to the impracticalities of size, heat, power, production etc..

I think a Cell derivative would be a perfect candidate for a handheld device and i'm sure with time, alot of the fruits of PS3 linux coding could really translate well to such a platform should Sony ever decide they want to go that route with the PSP2 perhaps..

 

[ebola]

heh yeah I found that thread later. Yeah defered rendering with tiled textures caching just like anything else is what I was thinking.

A psp2 featuring s/w rendering on the cell... hmmm so it may yet happen there is hope!!

 

[Lazy8s]

Due to the larger number of talented companies competing and the tighter size and power consumption constraints, portable GPUs are designed even more competitively than their desktop counterparts for performance.

The disparity between CELL and a customized design would only be greater.

 

[Blazkowicz]

software rendering on cell should be particularly interesting for linux on PS3, I'd like to see some Free games and demos with weird tricks, features and effects even if the texture filtering might have to not be great (something custom and weird like software Unreal engine 1)

 

[archangelmorph]

<removed>

 

[Shifty Geezer]

The disparity between CELL and a customized design would only be greater.

It's a cert' that Cell vs. handheld GPU would lie very much in the GPU's favour, but at the same time Cell would bring to a handheld a greater flexibility. eg. When not playing games, one SPE could be used for rendering a vector interface and the others could do voice recognition or speech synthesis. Cell could also handle media stream decoding, and be scalable for different products. The upside would be a powerful, versatile device which could not be easily matched with conventional ICs. With this as a target, an efficient Cell-specific renderer seems like a good idea, even if peak performance is below what the best GPUs can handle - the modern and future feature-rich media-phones will spend less time rendering 3D graphics than providing other services, so a 'GPU' that can aid these other functions has to be smart move.

 

[Lazy8s]

Power consumption is probably the most limiting factor in portable processor design, and sub-$20 portable SoCs like Texas Instruments' OMAP already have custom DSPs for multimedia and codec processing, already have graphics cores for vector graphics and shader flexibility beyond SM3, and already toggle power consumption among parts of the chip as needed for power savings.

 

[ebola]

With this as a target, an efficient Cell-specific renderer seems like a good idea, even if peak performance is below what the best GPUs can handle - the modern and future feature-rich media-phones will spend less time rendering 3D graphics than providing other services, so a 'GPU' that can aid these other functions has to be smart move.

I read about some DSP (i forgot the name) which did at least have a Lerped 'load' instruction (i.e. combine 3 of those to do bilinear filtering). Elsewhere on these forums i read the suggestion to add barycentric coord instructions... I bet (ok, hope) with a few simple enhancements you could get primarily cell driven rendering to work