Fafalada said:
Aside for both running GPUs on roughly same clock speed, what other confirmed chipset similarities are there?
Low main RAM latency: we go from Direct RAMBUS with embedded memory controller to mobile DDR RAM with embedded memory controller.
Simplified system to maximize real-world performance: single user-programamble CPU with easy to access co-processors (no need to handle DMA chains for VIFs and write and upload micro-code, handle double-buffering of inputs and outputs). The DMAC is also partially abstracted IIRC through the PSP System API and has a simplified job to do (with a nice new feature added like display list culling: improoves efficiency).
T&L, clipping, multi-texturing (I think it is still multi-pass, but that would be more evident when SCE will start to talk about using the VFPU in parallel to the GPU's HW T&L engine) all on the GPU.
GPU's TANDL engine is a fixed-function hardware accellerated pipeline like Flipper's T&L engine (although with much better support for Skinning [more bones IIRC]) and for custom T&L and deformable non static meshes you fall back on the CPU's VFPU.
No no, I meant your recent negativity about some PSP aspects and making conclusions based on DM(tm) logic.
Well, everyone does it... I wanted to try DM(tm) logic too
.
Seriously, one thing is to say that the GS is PSX GPU SLI (which is not true although I think they changed some things in the GS to ease backward-compatibility: somehow, I might be wrong of course, I do not think you would have to send verties in fixed-point format to the GS ) and one thing is to say that the PSP chipset can trace some of its root on the now abandoned GS2 track and on the EE.
In the GS2 SCE was already working on a refinement of the GS core with new features and performance improovements: it would still bear similarities, many, to the GS as the GS3 was supposed to be the brand-new GPU, not the GS2.
My speculation was that the PSP GPU rendering core was a cut-back (we know it has 4 pipes while the GS2 likely would have stayed with 16 pixel pipelines) GS2 or GS2 evolved (PSP devewlopment picked up where GS2 work ended in regard to the GPU rendering core maybe).
I was just getting the bad feeling that texturing would face the same problem as it did on the GS: single-cycle bi-linear required to use half of the Pixel Engines as TMUs basically halving the peak fill-rate (from 8x2 to 8x1 in regards to pixel-processing per cycle).
In the PSP case a textured fill-rate of 332 MPixels/s with plain bi-linear seemed a bit low (conidering the PSP is not a deferred renderer: I hope early-Z is hidden in there though) as if things were again the same as on the GS tri-linear wopuld have brought you down to 166 MPixels/s.
PlayStation 3 is where all the next-generation effort is: PSP was meant to be comparable to PlayStation 2 in terms of visuals, not to spank it silly.
PSP had to be about as powerful as PlayStation 2 in real-world scenarios and had to eb a low power design (in terms of power consumption and battery life).
Increased efficiency is one of the things you want to work on to meet both goals: it is what they did IMHO. Made the architecture more forgiving and with a less steep learning curve and a system that adapted better to a high level API similar to OpenGL.
With that said, why not leverage the R&D done for the GS 1.5 and GS2 (both had work started on them) and salvage and improove other things (such as the Macro Mode instruction set for the VUs [you know my theory on the VFPU]) ?