Perhaps I'm looking at the 4cm fan and 45nm things a bit too hard, but given just how low I think Nintendo need the BoM to be and how limited the cooling will be the be I'm just not expecting that much. I think it can get into the same ballpark as the 360 while drawing substantially less power because they should be able to save loads of power on the CPU by clocking lower.
-------------
BTW I'm expecting lower peak flops from the WiiU CPU, but that it's more resistant to code that makes the 360 bum out (OoOE, bigger L2, possibly higher L2 bandwidth, lower latency memory access etc). Super optimised 360 stuff might pose a problem though I guess.
-------------
At the time I thought the tech demos looked nice but couldn't see what was "next level" about them. Screen grabs showed that Link in the Zelda demo was actually really simple and low detail - way, way below a Mass Effect character, for instance. It's also possible that early dev kits may have been more powerful in some ways than the final system. As with the Xbox 360, perhaps final clocks won't quite live up to early expectations.
-------------
Maybe it won't be as easy to just throw a PS360 game at the WiiU and have it gobble it up as people were expecting, but with sufficient work the ports will probably be fine. I guess that could give put off some publishers that aren't expecting great market penetration with the hardcore.
-------------
I think it would be fun to see what people's lowball expectations for the WiiU are. At the very outside I think it might be possible for Nintendo to get away with a SoC with something like 160 shaders, 12 TMUs and 8 ROPS - at something like 700 mHz that would fit the rumour of "fewer shaders" while still giving you about the same minimum performance. I'm hoping for something better though.
I think that you are right (after reading the rumors about Nintendo being out for costs reduction).
For the GPU I pretty much agreed with Rangers that a down clocked RV730 would do the trick.
Actually more half a RV740 than a RV730 as the latter is a bit "strange" (8 40 wide SIMD vs the usual 80, with a lot of texturing power vs latter architecture).
I would discard GDDR5 as it is 3 times more expansive than DDR3, at least in large quantity. So edram may have a role to play in rendering (more on that latter).
Overall I would agree with you 2 SIMD may be enough still if it's a SOC and assuming low power CPU cores (4 power a2 are ~9Watts and CPU issued from IBM embedded parts should be in the same ball park or lower) I would think that 700MHz is too high. Low power llano SKU (using better process) are clocked in the 450MHz range.
But overall along with the fact that specs seems to be a moving target I'm close to dismissing a SoC all together. SoC are complex design there is not playing with the specs that much.
Accordingly to AMD own numbers the HD6450 has BOM of 38$ (I believe that the version with 512 MB of DDR3). It's a tiny chip 67 sq. mm
I can't see Nintendo going with a chip (for me 45nm one / speaking of the CPU) big enough so it could embark enough edram for make it relevant for rendering. Not to mention to complication of having the the GPU to access this edram.
As I see it the CPU chip could cost as much if not more than the GPU (I know that AMD price is for the whole GPU including RAM). At this point I would assume that the cpu will be less than 100 sq. mm too. I would (again) discard edram as being relevant to rendering.
So putting altogether, I could definitely see Nintendo shipping something like this:
CPU:
Tri cores from the embedded line.
OoO, no SMT
2MB of edram (the L2 cache basically)
CPU speed could be set anywhere between say 1.5 and 2Ghz
64bit bus to either 512MB of ram.
RAM would be DDR3, clock speed anywhere between 533 and 800 MHz
Bandwidth to the main ram would be anywhere between 8.5GB/s and 12.8 GB/s
between 10 and 15 watts
(I'm using
the info from this page because I believe that if that kind of budget ram is relevant to AMD low end gpu it is to Nintendo too).
GPU:
Caicos/ HD 6450 / HD 6400M
160 Stream Processing Units
8 Texture Units
16 Z/Stencil ROP Units
4 Color ROP Units.
64bit bus to 256MB of VRAM
RAM would be GDDR5, clock speed 800Mhz.
Bandwidth to the VRAM would be 25.6GB/s
I'm using AMD own data (the same as above) and I chose the slowest option for the gddr5 (for price and mem controller power consumption).
Then there is the clock speed.
Anand gives the hd6450 TDP @750Mhz with 900MHz GDDR5 at 27Watts. It's too high. So I already cut the ram speed I would also cut the chip clock speed.
I may put the clock around 500Mhz (it could end lower).
I would not be surpised if the power budget for the GPU is in between 15 and 20 Watts. At 650 MHz the hd5450 is almost 20Watts and that's with DDR3. So taking in account the power cost of GDDR5, sadly ~500Mhz sounds right. That's 160 MFLOPS
Both chips could be connected by a pci express x8 link.
------------------------
That's pretty much it, Edram is too complicated for rendering especially as the wiiumote ups the requirement for the framebuffer (two of them screen and WiiU).
I know fans will try to kill me but I believe that it makes a lot of sense, 256MB may make up for the higher cost of gddr5 vs ddr3 and Nintendo needs that bandwidth.
512MB freed from the FB requirement should be an improvement over the ps3.
The system would consist of two tiny and cool chips, max tdp would be 35Watts. Passive cooling and a unique fan in the box for the power supply, the chips, etc. should do the trick.
The cost using numbers given by AMD and extrapolating for the CPU north of 80$ including the memory chips.
Now I'm ready to face the fans hatred
EDIT
For the sake of beyond3 article I may push the number of SIMD to three put it would be "much ado about nothing" to go custom vs off the selves part for so low gains.