Wii U hardware discussion and investigation *rename

Nintendo might need the memory for its miiverse feature. In a way it's like PS Home done right. It provides a persistent online community with different titles having their own hubs, and unlike PS Home you can access it instantly from any game.

The 480-576 gflops figure comes from the 4670 right? 1.5x the clockspeed of xenos and 1.33x the ALUs. A downclocked 4850 at 500Mhz would be about 800 gflops.

 

Is a 4670 an RV730 I assume? Well that's what I've assumed all along.

An 4850 would be a major increase in die area.

RV730 vs RV770

956m transistors vs 514m

Die size, 146mm vs 255mm (on 55nm)

Yeah, my only trepidation is if Nintendo would even go as beefy as RV730. There is absolutely no way I see them using the RV770.

Hmm, the e6760 bg touts, I see where he got it now. It clocks exactly 576 gflops. It features 480 SP's, so it's weak enough I could believe it. It does use GDDR5 though, which means they would have had to rework the bus. Regardless, just because it could have been in a dev kit doesn't mean it's in final.

http://www.anandtech.com/show/4307/amd-launches-radeon-e6760

It has 716m transistors and would be DX11. I dont think that's it, I think it's an RV730. I just question why Nintendo would spend 700m+ transistors on something all the devs are saying is in the same ballpark as PS360. This could be because they severely downclock it, but again it's wasteful.

The 4670/rv730 clocks 480 gflops at stock 750 mhz. Assuming they downclock it significantly from there, you can sensibly get in the range of better than Xenos/RSX but not mightily so. It also natively takes DDR3. Not being based on a mobile part should make yields better.

 

Yes, we've been told as much. But we don't know to what degree. Nintendo chose it as they got a good value proposition for their console - the same reason any company chooses any component (other than scrambling for a last minute part ). We don't know any of the details of the chip so can't explore specifics.

Watson featured on a TV programme and then got coverage in the press. While Blue Gene hasn't had that same mainstream coverage AFAIK, and isn't POWER based anyway, so why would they talk about it?

It's PR. All PR is crap! That was copy written to explain a very dull photo of a memory module. It is clearly written by someone with no clue as to the tech involved in consoles - IBM don't get their engineers updating their website's press room AFAIK.

There are lot of points that can be considered here, not least of which if a tsk is running in the background, what do you care about performance? But remember this processor is going in a home console being sold on it's gaming function. Nintendo have never talked about multitasking OS features. Lots of eDRAM L3 is a very expensive addition to a CPU. If the benefits are mostly multitasking, how much are Nintendo really going to want that?

16 MBs, and it's still damned expensive.

It'd be ideal for the GPU to have access, but I believe that'd require a very different memory architecture. If the GPU is accessing the eDRAM on another chip, than it's not embedded to the GPU so it won't gain any advantage. If the eDRAM sits on a die with both CPU and GPU attached, then it could serve both, although the CPU would probably just get in the way of the GPU's memory operations.

It's a custom part and not an off-the-shelf POWER7.

We're likely looking at something more like Xenon. Maybe it's a tiny chip so that it can be included on the GPU die, using IBMs eDRAM with the GPU??

 

disappointing news
http://spong.com/feature/10110776/Preview-Sonic-All-Stars-Racing-Transformed

 

I think the key word here is, release "window".
That can be up into March, for all we know.

But its strange when...

http://www.nintendolife.com/news/2012/08/sumo_digital_wii_u_looks_as_good_as_hd_platforms

 

Here is the thing, when I saw what they did with the Xbox Slim, I figured this
was basically the direction that Nintendo was going for:

(And this is purposefully bottlenecked)

If Nintendo is going the same direction, I would just like to see how Nintendo/IBM can fit Power7 cores alongside the AMD GPU. Maybe somebody here can visualize chip placement on 45nm silicon.



That L2 Cache, could that be the eDRAM for the WiiU?

 

Actually, the DX10.1 R700 generation of AMD GPUs has already compute shader support (and nVidias DX10 GPUs are arguable even better in this respect). DirectX CS with shader model 4 directly maps to the features offered by this AMD generation (OpenCL has worse support [makes no use of the more limited shared memory, DX CS does]).

 

The Wii U CPU cores could be anything. We've no idea what size they'll be, so can't envisage any integrated package. We don't even know the size of the GPU, although could hazard a reasonable guess at this point.

Well, that's an XB360 die. Might well bare no resemblance whatsoever to Wii. eDRAM would be pretty slow for an L2 cache, but maybe larger cache as an L2 would be beneficial? You'd need a CPU engineer to tell you that one.

 

Yeah, it would be slower, but the power savings and are might be worth it might. IBM PowerA2 uses eDRAM for the L2 cache. Slide 13 from the link below says that using eDRAM results in about 3x cycle penalty and 2.5x latency, but they came application performance is similar in about ~50 of the area and ~20 of the power.

With some developers complaining about the performance of the CPU, I wonder if it because they're used to high throughput vector units (360) or SPE's. In the Eurogamer article, the one dev that didn't have issues was one that didn't rely on heavy CPU use. That to me point suggests a weaker vector unit (due either to width or just the lower clock) in the WiiU.

If I had to guess about the WiiU CPU, I'd say the Peformance/Watt will be much, much better than the other two consoles, but I wouldn't find lower Raw performance (gigaflops) to be surprising.



https://www.google.com/url?sa=t&rct...6LnaYnLU1NKQ17_Kg&sig2=-EJD2hwnwIXER6Dkn3bFqQ

 

Bah. You're twisting things on me again.

First the original dev kit was ~576GF. I'm expecting the final to be over 600 so there's no need say it's "bg's 576GF". The only reason I've "touted" the E6760 is because of its TDP and performance and that I see. It's GF was more of an interesting

Also the 4770 is 137mm @ 40nm, and I'm guessing/expecting the final Wii U GPU to be on a 32nm process.

What you're talking about isn't the same thing. To me it sounds like you're talking about general GPGPU capabilities. Compute Shaders were added first to DX11 and I just learned OpenGL 4.3, which was released a few weeks ago, has now added it. So either Nintendo has no clue what they are talking about, or there is merit to that being specifically mentioned.

 

The RV7x0 architecture has the set of hardware capabilities required to run standardized Compute Shaders equivalent to what can be found in the current DX11 and the latest OGL APIs.

Given that the Wii U operate in a closed environment, standard APIs level of requirements are not an issue here. Simply put, a RV7x0 based Wii U GPU wouldn't need extra logic implemented to support Compute Shaders.

Edit: To be more thorough, you can check this MSDN article explaining the API differences. Although it's not really relevant to the specific API Wii U uses, it can give you an overview of what to expect to be the differences between the various Compute Shaders versions available:
http://msdn.microsoft.com/en-us/library/windows/desktop/ff476331(v=vs.85).aspx

 

OpenCL has been around for a lot longer though.
Probably Nintendo would obtain ATI's existing compiler as a starting point.
The compute thing though is complicated, having compute and making it useful are two entirely different things.
Outside of the difficulty writing efficient compute shaders, and the efficiency they run on any given piece of hardware, devs will think long and hard before stealing GPU resources from rendering.

 

What exactly are you disagreeing with that I said about you?

Apparently I'm twisting things but everything I said is exactly right according to your own next two sentences. OK. Awesome.

You need to stop getting upset over nothing. There's nothing wrong with me staying you touted something when well, YOU DID. By your own admission! You've thought the E6760 was the most likely candidate, or at least it's the one you mention most, for a while. What's the problem?

 

Well if its rendering on the tablet at the same time.. that might easily eat up the GPU edge plus APIs and tools need a few years to be comparable

 

It's only rendering a map on the Wuupad screen. Sounds more like optimisation woes and probably will be fixed come release.

 

I didn't know that. I think that's our most likely candidate there - the eDRAM is L2, and the chip is a very custom part like Xenon, only perhaps without the vector throughput. Probably low power as you say, which fits the thermal constraints many expect of the small form-factor.

 

Thanks for the link. I've read some things similar to that in the past though and this is why I said it would be "weird/wasteful" as the capability for 10.1-level hardware to have CSs is through the DX11 API. So I guess to elaborate a little more, it would seem strange to me that Nintendo would use less efficient hardware if they plan to implement something like that.

That was my point in the first post of mine talking about it being "weird/wasteful".

And right I'm familiar (enough) with OpenCL and ATi Stream. But Compute Shaders as I've seen it wasn't introduced till DX11.

Haha. I don't get upset over stuff like this. When I post I am direct and emotionless, and that ends up being open for people to interpret however. That's why I've developed a habit of using emoticons more than most do to help "soften" my posts.

I put tout in quotes for a reason. In other words you're saying I'm giving it way more weight than I really have. I've never said it's the most likely candidate. My post history is free to be searched and studied if you want to. I know what I've said in regards to that so I feel I have no need to defend myself at this time on it.

 

And you can run them without problems on DX10 or DX10.1 level hardware, if it supports CS4.0 or CS4.1 (and the compute shader is written within the restrictions of that shader model). All nVidia GPUs starting with G80 support CS4.0 and all AMD GPUs starting with RV7xx support CS 4.1. So if you install the DX11 runtime, you can run DX ComputeShaders on DX10 or DX10.1 level hardware without any issues. CS5.0 support starting with Evergreen and Fermi basically just loosens some restrictions and adds a few features. But it doesn't mean one can't write and run useful compute shaders for DX10(.1) level hardware. It's possible within the DX framework and it would be possible within the Wii U's proprietary API. And it's definitely enough to put it on a list of bullet points.

 

Serious question to bg: you expect to find 4850 PC level performance in this box in console environment? I must have missed something

 

Issues weren't my concern though. Efficiency is. After all the way I see it if it didn't work well, MS would have swept it under the rug and pretended like it never happened. Everything I've seen in regards to that suggests that level of hardware is not as efficient due to lacking certain features. If you have have something that shows otherwise, I'm willing to back away from that belief.

Only if the GPU shapes up like I think, and even then that's relying on the Vantage P score as the determinant to that comparison. And that ignores possible usage of general processing on the GPU in some games.