CELL as GPU did make out of theory land as SPUs on Cell were constantly used for graphical tasks. In fact you can fire up youtube and look for raytracing stuff which was entirely done on CELL SPEs.
They stuff you are talking about like ROPS are backend that you could just attach to silicon or get it as separate silicon if needed and not that important.
The fact that CELL could be used for GPU like tasks and software rendering was not news, I am aware it was used for these and many more tasks (DICE had a clever way to use them in frostbite to help with geometry culling too)... any CPU could and if you take programmable and self-feeding wide vector processors with their own local storage and huge register file are of course a great candidate for such work... they sounds awfully close to any modern shading core
.
What they lacked, and what the patent detailed (and what LRB added on top of their vector unit too) was fixed function hardware to fetch and filter textures, rasterize triangles, merge MSAA samples, blend render targets, etc...
There were 2 reasons why double CELL which was initial design didn't work out.
1. Cost. CELL turned out to be way more expensive that they predicted and moreover late. One Cell in PS3 already placed them in red so two of them would quickly sink Sony.
2. Paradigm shifts in how graphics were made.
When CELL was designed most of graphics creation was still not standardized. Back in PS2 days you had 2 VPUs which were precursors to SPEs in CELL and they were the one doing heavy lifting when it comes to graphics.
Funny thing is that the VU's in PS2 were much more similar to the capabilities of modern shaders (much more flexible than vertex shaders of that time)... but that is just an aside and to be frank not to dissimilar to the VLIW based shader cores you can find in older Radeon GPU's (think pre-GCN).
So CELL was natural progression on that front. Instead of 2 VPUs you would get ultra fast 8 SPEs per CELL.
It was perfectly reasonable for Ken Kutaragi to assume now legendary 120FPS for every PS3 game if they based their prediction on what could you do with PS2 level of graphical technology with PS3 hardware.
[...]
So they had to throw out one CELL and get GPU that can do shading well. Nvidia chip alone also was actually behind paradigm because in shading world texture and vertex shading was combined by AMD while Nvidia GPU for PS3 still operated with separate shading modules so from get go it was worse than Xbox gpu.
And that is how PS3 was made. If we would still operate with PS2 graphics technology world then PS3 would absolutely murder competition graphically.
I am aware of the many reasons why CELL based GPU's did not fly, despite being a lot of interest in the field much after CELL was launched and ended up not flying as high and as wide as expected. Also aware that the flexibility of having customised universal shader units provides maybe less peak performance in some tasks, but overall the flexibility it provides to developers (being able to dedicate all units to vertex shading or pixel shading depending on the workload is a massive boost... think z pre-pass)
The LRB project had the same basic philosophy: since GPU's are becoming more and more programmable and that is where they believe the next leap is, let's find the minimum possible fixed function HW that can take a sea of optimised x86/general purpose cores so that they become efficient at pursuing graphics tasks against modern GPU's while allowing developers with exploring the flexibility that software rendering allows.
There is a famous old Sweeney interview were he was also highly anticipating CPU's becoming fast enough at such massively parallel tasks to allow rendering to move back or mostly back to software.
The paradigm shift came about 2002-2004. Shading was introduced as standard in games with Doom 3 being harbinger of things to come. Soon texture and vertexes became unimportant and shading power became main resource hog.
And this is something CELL couldn't do well because it was designed with old paradigm in mind.
If anything SPE's would better at purely shading tasks then they would in a "texture based world" as they lack dedicated HW to process textures. The reason they lost and LRB lost too is that the performance advantage, developers tooling / programming model available on GPU's, and power consumption and silicon cost of the dedicated HW surrounding and running the shader cores was too great and is still too great to replace them with general purpose CPU's even with augmentations.