Still though can any compression out there make up for the 256 gig buss between the two.
Do the math and you tell me
Still though 30MB of edram would fit 720p 4x fsaa and 10bit HDR .
Yet what if you want 1080p @4xMSAA, A G-buffer and A-buffer, FP16, a deeper Z-buffer, and a host of render targets? Now you are back to many, many tiles and we already know what developers think of this approach.
Essentially with eDRAM you are dictating to one degree or another what resolution/features as well as rendering approach you take.
Current gpus have large amounts of fast ram.
They have large amounts of fast memory for a reason. The question where the demand for that bandwidth is being spent. Is 30MB of eDRAM to be stuck at 720p 4xMSAA (or 1080p 2xMSAA) and 32bit color and a huge PITA to deploy a number of emerging rendering techniques worth isolating fillrate bandwidth from the rest of the system?
How much is your fillrate cost going up over the next 4 years? How much everything else? What is your projected bottlenecks? GPUs already have a lot of overdraw ability, so is the silicon cost and associated limitations from a software perspective worth it?
next gen consoles will most likely be a 128bit bus again.
The PS3 isn't using a 128bit bus--it is using 2x128bit busses (XDR, GDDR3). The Xbox 360 has a 128bit bus to GDDR3 plus the daughter die bus. So neither console is simply using "a" 128bit bus. Further, while pad size reduction is much slower than process node reduction it does progress. Some of GPUs with 256bit buses to GDDR5 are relatively small as well as have robust bandwidth--enough bandwidth to push 1080p with MSAA as I noted above. Hence back to my point: is eDRAM a good tradeoff? Is fillrate going to be a bottleneck next generation OR is the cost of a 30MB eDRAM (270M transistors) going to be better spent on ~ 400GFLOPs+ worth of vector units? We already established 30MB of eDRAM is quite small, so you are now talking about a reduction in rendering techniques as well as a hefty percentage of rendering performance.
Technology doesn't progress on a linear path, neither do rendering costs. We already saw in the GPU arena how ATI came out with a 512bit bus (R600) but has transitioned back to 256bit busses with the emergence of GDDR5.
I also doubt we will see anyhting but a uma memory set up next gen.
It will be interesting to see how they do an UMA with the trend toward CPUs with integrated memory controllers. I wouldn't say the NUMA is dead just yet, only that the tradeoffs need to be justified elsewhere.
Which means they will need a very fast pool of ram some where for the gpu. no matter how much smarter they make the gpus.
If they go UMA, why shouldn't this be a large pool of GDDR5 or even XDR2? I believe Hynix has already touted 160GB/s on a 256bit bus via GDDR5. Qimonda announced volume production up to 4.5Gb/s GDDR5 (the Radeon HD4870 uses 3.6Gb/s modules; 115GB/s over a 256bit bus). If we are looking at a 2011 window or later there is always the potential for other memory technology to be deployed as well. But the point being, still, is that games like Crysis (which has a lot of fillrate demands) isn't being crippled by having less theoretical bandwidth than Xenos--i.e. fillrate isn't a primary bottleneck to achieving those level of graphics. Not to mention there is a strong arguement that the eDRAM is much about cost and giving full fillrate to the 8ROPs than the technical bandwidth which should be balanced out by the 32GB/s bus between the GPU and daughter die. The 256GB/s of bandwidth on the daughter die isn't the same thing as having 256GB/s of graphics bandwidth.
Anyhow, memory management and design is going to be a major hurdle next gen due to the growing needs of multi-core processors and evolving GPU designs. From this perspective the idea of a simple design like Larrabee (or whatever NV/AMD pitch in the same mold) may be where the simplification comes and not necessarily in eDRAM.
