This is getting more ridiculous with every next post..
This is what i said. Lets quote the entire paragraph rather then your little selective quote.
Indeed, let's.
zzz said:
Given the Wii U's slow MEM2 bandwidth, it's also a fair assumption to say it's unlikely the GPU has more then 8 ROPs. ROPs are heavily dependant on bandwidth, there would be absolutely no point going with any more then 8 ROPs on a MEM2 bus as slow as the Wii U's. The Xbox 360's architecture had the ROPs tied to its eDRAM via a high speed bus, that also doesn't seem to be the case with the Wii U. The Wii U's eDRAM seems to be implamented differently and not tied into the ROPs, nor is the Wii U's eDRAM capable of offering bandwidth in the same ball park as that in the Xbox 360. If anything the Wii U's ROPs may be worse in performance then those in the Xbox 360 due to the shit bandwidth. Either way 8 ROPs for a modern day console is terrible.
Is that line i've underlined wrong?
Yes, it is wrong.
You have two memory pools A and B available to a GPU. A being of known parameters (size, bandwidth, etc), B - an unknown, outside of its size. ROPs work with pool B, while both pools A and B service asset fetches. Any statement, 'Pool A has such and such BW characteristics, so that ROP numbers should be such and such.' is nonsense - you have neither the BW available to ROPs alone, nor the cumulative bandwidth available for GPU assets.
Here's a hypothetical question for you: what if pool A was not present (or GPU could not directly read assets from it), how many ROPs should a GPU outputting to pool B have? 0?
And here's another one, this time non-hypothetical: how many ROPs do you think PS2's GS had, given its link to the main mem was rated at 1.2GB/s? GS has a top fillrate of 1.2GPix/s when texturing (2.4GPix/s sans texturing), compared to Xenos' 4GPix/s. By your reasoning, GS should have had 22.4 * 1.2 / 4 = 6.72GB/s BW from main mem for GS to have as many as 8 ROPs. Now, humor me and check the actual specs.
Does that sentence state the 8 ROPs are directly and physically tied in to the MEM2 bus?
No it doesn't.
So '8 ROPs on a MEM2 bus' in that underlined passage is supposed to mean what exactly?
Thus why i then went on to talk about the eDRAM in the XBox 360 and its relationship with ROPs and compared that to what i believe the Wii U's eDRAM is structured like. As it's obvious to even blind freedy 12.8 gigabytes per second isn't going to be enough. Even with the eDRAM, for certain ROP functions you're going to have to use the main memory. With MEM2 bandwidth as slow as the Wii Us, ROPs are going to be very limited in performance.
What are those 'certain ROP functions' you need main mem for?
In short piss of with your selective quoting and intentional misinterpretation of my post.
Edit: Also you are a fool if you believe the Xbox 360's ROPs didn't use its GDDR3 memory pool for ROP related tasks. IT DID. The eDRAM in the Xbox 360 wasn't sufficent to store every thing you'd need to feed the ROPs.
http://www.beyond3d.com/content/articles/4/3
Which again highlights why the Wii U's MEM2 bandwidth is important even for ROP performance. And why the low bandwidth of the Wii U's MEM2 pool strongly suggests this console has similar ROP performance to the Xbox 360. MEM2 performance is still a limiting factor in ROP performance.
In what way, form or manner do you believe that article supports you claims? Apropos, 360's GDDR3 is the
sole pool capable of feeding assets to Xenos (locked L2 non-withstanding). That means that every single render target (RTT) had to be first resolved to main mem, and then read back from there. In contrast, U-GPU is able to skip the entire roundtrip to main mem for all cases where the former and current rendertargets fit in eDRAM. Furthermore, Xenos' 256GB/s of ROP-only BW served one purpose and one alone - to provide free MSAAx4 with alpha blending (which, apropos, went down the drain the moment tiling was involved). Its actual eDRAM bus BW of 32GB/s serves perfectly well the 4GPix/s of max color fillrate that Xenos actually had, sans blending. For blended color pixels sans MSAA, a 64GB/s BW would've sufficed.
edit: Ok, I did neglect the 32GSample/s of zexel performance, which too, is enabled by the 256GB/s BW, but that has little to do with main mem BW.