Wii U hardware discussion and investigation *rename

Status
Not open for further replies.
So someone brought up the GF116 in another thread, and boy does it have a crazy memory setup (1GB on 192-bit ):

http://www.anandtech.com/show/4221/nvidias-gtx-550-ti-coming-up-short-at-150/2

It got me to thinking about the 768MB, >1GB, and 1.5GB rumours. Naturally, to hit 1.5GB, you need 12Gbit worth of chips. The easiest path whilst maintaining a single density is 6x2Gbit or 3x4Gbit. Given the limited motherboard space, I'm inclined to think no more than 3 chips per side (on top and underneath).

For 768, you just halve the density or # of chips. 3x2Gbit or 6x1Gbit

Code: 
1+1, 1+1, 1+1 Gbit or 2, 2, 2 Gbit  ->  768MB
2+2, 2+2, 2+2 Gbit or 4, 4, 4 Gbit  -> 1536MB
And I thought I recall seeing some coy remark about >1GB but not confirming 1.5GB... so what if...

Code: 
1+2, 1+2, 1+2 Gbit -> 1152MB
>_>
<_<
ಠ_ಠ

Bus width would be another question entirely (96-bit or 192-bit) as well as the RAM type. Both GDDR5 and DDR3 have up to 2Gbit densities available whilst DDR3 alone has 4Gbit and up (from what I can google).

------------------

A hypothetical discussion at Nintendo might have been...

"768MB?"
"Maybe too little... not much of a leap over 360"
"But dev kits traditionally have double the RAM"
"Yes, but we painted ourselves into a corner with the chassis size"
"Multiple Densities!"
 
http://uk.ign.com/articles/2012/01/24/xbox-720-will-be-six-times-as-powerful-as-current-gen


IGN says X720 will have the equivalent of a HD6670, which is 20% faster than the Wii U.
If X720 has the equivalent of a Turks (Northern Islands 480sp, 24 TMU, 8 ROP), then the Wii U could have something along the lines of a Redwood (Evergreen 400sp, 20 TMU, 8ROP).


They also claim this HD6670 translates into 6x the graphics capabilities of Xenos. I honestly don't know how a 718M transistor chip at 750MHz can be considered be 6x faster than a ~350M one at 500MHz. Even with the generational gap considered, there's an overhead related to DX11 compliance.
Plus, this 6x performance advantage doesn't make any sense while looking at the fillrate and flops from each GPU.



Of course, this is IGN talking about GPUs, which is about the same as a monkey talking about quantum physics... so there's this chance that although their sources are right, this guy might've heard "sixty-eight hundred" and at the time of righting thought "was it sixty-eight or sixty-six? bleh I don't care, it's probably the same thing".


For example, a Barts-based GPU (Northern Islands, 960/1120 sp, 48/56 TMU, 32 ROP) would fit that performance description quite nicely, with the Wii U getting something along the lines of a Juniper (Evergreen 800 sp, 40 TMUs, 16 ROP).



Wrapping things up, either the "6x" claim is false or the "HD6670 performance" is false, or both are false.
 
Develop is reporting that Wii U is twice as powerful as Xbox 360

Nintendo’s next generation hardware will be roughly twice as powerful as Microsoft’s current system, the Xbox 360, according to a studio source speaking to Develop.

The person, communicating anonymously from a studio currently building a Wii U title, said the new Nintendo console could achieve roughly twice the processing and graphical potential of Microsoft’s current generation machine.

While twice the power of an Xbox 360 is broadly above market expectations, Develop’s source claimed this is in fact less than some studios had expected.

“I've heard [a project designer] complain it's underpowered compared to what Nintendo announced, resulting in people having to de-scale their plans,” the source added.
Click to expand...

more here: http://www.develop-online.net/news/39593/Wii-U-twice-as-powerful-as-Xbox-360
 
I get the feeling that MS want to keep thermal specs similar to 360. Putting something with power usage similar to Cayman into a set top box might be problematic, especially considering how they had engineering problems with 360 already.
 
ToTTenTranz said:
They also claim this HD6670 translates into 6x the graphics capabilities of Xenos. I honestly don't know how a 718M transistor chip at 750MHz can be considered be 6x faster than a ~350M one at 500MHz. Even with the generational gap considered, there's an overhead related to DX11 compliance.
Plus, this 6x performance advantage doesn't make any sense while looking at the fillrate and flops from each GPU.
Click to expand...

6x is definately a stretch. 4-5x isn't unreasonable if you're looking at pure shader performance though. Its 3.3x faster on paper but efficienct gains between R500 and R9xx would add a lot of performance too. Just look at how much faster the 3870 was compared to the 2900. And the 2900 itself should be more efficient than Xenos.

For example, a Barts-based GPU (Northern Islands, 960/1120 sp, 48/56 TMU, 32 ROP) would fit that performance description quite nicely, with the Wii U getting something along the lines of a Juniper (Evergreen 800 sp, 40 TMUs, 16 ROP).
Click to expand...

Sounds about right but only if running pretty slow, say 500Mhz.
 
I mentioned it was ancient because been linked and discussed on this forum already.

And I think our definitions of amazing are quite different.
 
Kaotik said:
While it does add something, we do know it looks far better than the official footage we saw, and there's no direct feed available of the build on the show floor
Click to expand...

Yeah, for whatever reason, the direct feed release doesn't have as many effects as the full thing shown on the show floor. One of the most noticeable difference is the self shadowing that we don't see in the direct feed.
 
Teasy said:
Nobody said self shadowing is new technology..
Click to expand...

Considering the math and basics of programming self shadows and most now real time capable effects have has been around since the 1970s, I'd say it's old technology :p It's quite amazing what people at the University of Utah were doing back then.
 
-----------

Had a couple thoughts about the rv740/4830 rumour for the early WiiU dev kits. The mismatched rumours can be reconciled if you consider the mobility radeon 4830, which is actually based on rv740, and is thus a 40nm part. I'll note that the 4830 mobility was clocked at 400MHZ at the lowest and was paired up with either DDR3 or GDDR3 on a 128-bit bus.

That said, if they were already having heat issues, the only logical path would be to hope for 28nm, lest they gimp the GPU even further. I'm still not convinced that Nintendo is going to suddenly absorb the high costs of securing 28nm capacity this year (i.e. 1 million GPUs for Q4 2012). But one can always hope for it since that's the natural course to take, and I have no doubt that was the intent, but being beholden to TSMC's schedule (or Global Foundries for that matter) was always going to be precarious, which makes this 2012 launch that much sillier in the grand scheme as far as technology goes.

There's still a problem regarding the size of the chip in that instance. rv740 is a 137mm^2 chip and moving to 28nm would probably make that too small to use a 128-bit bus, let alone a 192-bit one. Of course, that's precluding the event that AMD beefs up the chip even further so that it's just big enough.


--------------

Of course, if what was seen was a "big" chip, then that doesn't jive with the 140mm^2 chip, which is for all intents and purposes, pretty small in its own right. The reasoning put forth thus far is too vague for anything to be concluded ("seeing a big chip, therefore process node is certain"). A big chip would be rv770, the original 256mm^2 die, where the 4830 has two disabled SIMDs and is thus equatable to the rv740, albeit with differing process nodes. Given the vagueness so far of just what that size is, there's nothing definitive about the progression from here. If the original kits stemmed from the 55nm line, then 40nm would be the logical progression.

As it stands, we have nothing definitive about the particular technology used in the kits or if all the kits were the same or how the kits were updated.

The reports of the WiiU being underpowered doesn't really work considering just how much power a 4830 mobility has over the current gen, even if it were clocked as low. There's little to reason to suggest that it's low power, unless the developer who made the comment was only expecting more (for whatever reason).

-------------

If we also consider that the thing is supposed to have eDRAM strapped onto it, we can reason that the I/O bus is going to be fairly wide on top of a 128-bit bus, so that already places a minimum die size that's possible (north of 100mm^2 if we consider a similar I/O footprint as 360's GDDR3+eDRAM buses). So in that sense, a straight die-shrink of an rv740 would never make sense, it'd have to be bigger.
 
Status
Not open for further replies.

Similar threads

Rys
Reopening Architecture and Products + rules and guidelines changes
3 4 5
Replies
90
Views
16K
Shifty Geezer
Shifty Geezer
Back
Top