WiiU "Latte" GPU Die Photo - GPU Feature Set And Power Analysis

Status
Not open for further replies.
...(snipped for brevity)...

Traditionally (as in Flipper & Xenos), embedded console framebuffers have held only the back buffer - all the front buffers sit in main RAM, at the expense of the BW required to copy a fb out of edram. I don't know how Latte handles that - it could very well be a matter of developer's choice.

Would this explain why most ports have been quite awful?

Switching a buffer to the DDR3 RAM and not using the larger space available on the WiiU's eDRAM because it's easier/has been commonplace to do so?

Cheers
 
I think the word 'awful' is a bit over the top, they haven't been that bad. The ports haven't been top-notch for several reasons - unfinished dev kits and tools, second tier developers for the most part and porting games from two 'CPU heavy' consoles with in-order and multithreading to a console that's 'GPU heavy' with out-of order processing and no multithreading.
 
Regarding the eDRAM bandwidth and the lower estimate, would this be unrealistic now that Shin'en have confirmed that both the GPU and CPU can access that 32MB..? It would have to be higher than 70Gb/s to prevent bottlenecks, wouldn't it..?

How do you mean? It may be like many console designs where the GPU gets fed the full bandwidth and after passing through that, the CPU gets a lower amount of bandwidth to it, like the PS4, One, 360, etc etc designs. In fact, it's certainly like that, looking at the layout.

Why do you say 70GB/s would make a bottleneck? How do we know that? And even if it was a bottleneck, there's no assurance that it's not true just because it's a bottleneck.

Saying it's more likely 130GB/s instead of 70 just based on which would be "better" is kind of unfounded. I'll have to look back through when the bandwidth was first debated, but I felt sure the 70GB/s was the far more likely position? It was based off the GPU clock speed of 529 or whatever MHz and the pin count on the internal eDRAM I believe. And in fact, looking back through the thread, it wasn't 130GB/s that was the possibility at all, it was 103.

Xenos runs its eDRAM at 2 Ghz to achieve that throughput to the ROPs. Wii U's eDRAM is likely running at the GPU clock although 800 Mhz is also a possibility so we're probably looking at 70.4 GB/s or 102.4 GB/s. It seems devs are choosing to take a post processing approach to AA and save the eDRAM for local render targets, framebuffers, scratchpad, and other duties.

Somewhere in this thread we did a number swap, whoops, lol. 103, not 130, and that was based on the assumption of the eDRAM being clocked higher than the GPU despite being in the same die.
 
How do you mean? It may be like many console designs where the GPU gets fed the full bandwidth and after passing through that, the CPU gets a lower amount of bandwidth to it, like the PS4, One, 360, etc etc designs. In fact, it's certainly like that, looking at the layout.

Why do you say 70GB/s would make a bottleneck? How do we know that? And even if it was a bottleneck, there's no assurance that it's not true just because it's a bottleneck.

Saying it's more likely 130GB/s instead of 70 just based on which would be "better" is kind of unfounded. I'll have to look back through when the bandwidth was first debated, but I felt sure the 70GB/s was the far more likely position? It was based off the GPU clock speed of 529 or whatever MHz and the pin count on the internal eDRAM I believe. And in fact, looking back through the thread, it wasn't 130GB/s that was the possibility at all, it was 103.



Somewhere in this thread we did a number swap, whoops, lol. 103, not 130, and that was based on the assumption of the eDRAM being clocked higher than the GPU despite being in the same die.

Is eDRAM clock based on Renesas numbers or is the clock based on assumption. Is there any reason why it wouldn't be clocked higher than Latte?
 
I know that, and that's why I speak only for the few that are doing the same that Krizzx did but on the opposite direction (mostly USC-fan and KidBeta).

So trying to have a discussion / asking for proof that the Wii U is based on something other then the R700 that is present in the dev docs and dev kits (iirc) and also trying to point that it is far more likely to be bottlenecked by other things to effectively use 1+TB/s of ram is doing with Krizzx? no, im sorry, thats crap.

I don't think the Wii U is a lemon, I just don't think that is 100% from the ground up Nintendo hardware which increases efficiency by large amounts like people in this thread where previously suggesting.
 
Nothing major it seems, but just wanted to note that a Gametrailers Review went up for Wind Waker HD. Apparently the "frequent frame drops are all too noticeable."

http://www.gametrailers.com/reviews/orhjbl/the-legend-of-zelda--the-wind-waker-hd-review

Considering it's a small project, I wouldn't be too worried though.

Maybe they're wind waker noobs and are talking about the effect when your weapon strikes an enemy and it feels like the game drops frames? At least it did to me many eons ago, before I realized it was intentional. Or maybe not. The footage I've seen hasn't shown any excessive frequent dropping though.

Edit: azak it's no longer really a GameCube game. There are numerous changes that require a lot more graphics processing than the original, not the least of which is a huge resolution jump.
 
Nothing major it seems, but just wanted to note that a Gametrailers Review went up for Wind Waker HD. Apparently the "frequent frame drops are all too noticeable."

http://www.gametrailers.com/reviews/orhjbl/the-legend-of-zelda--the-wind-waker-hd-review

Considering it's a small project, I wouldn't be too worried though.

Pretty terrible to be honest. For a GameCube game, of which the Wii U should outpower it beyond belief, to see framerate drops (Where I assume the original didn't have) is rather sad.
 
Maybe they're wind waker noobs and are talking about the effect when your weapon strikes an enemy and it feels like the game drops frames? At least it did to me many eons ago, before I realized it was intentional. Or maybe not. The footage I've seen hasn't shown any excessive frequent dropping though.
Maybe it's a data streaming thing causing hick ups?
 
Pretty terrible to be honest. For a GameCube game, of which the Wii U should outpower it beyond belief, to see framerate drops (Where I assume the original didn't have) is rather sad.

I assume the original did have them.

I assume StevieP is right and they don't know it's an animation of the game and not a drop.

I assume. The game.
 
Maybe it's a data streaming thing causing hick ups?

It's possible. Or it's possible they've overextended the GPU budget as well, doing all the new light rendering or post processing or something. Or maybe it's nothing. If someone remembers what the data rate off the GC disc was, feel free to chip in.
 
again one of the developers of the hardware at Nintendo said this.

I would draw attention to how efficient it is. For a computer to function efficiently, memory hierarchy structure is very important, and this time the basic memory hierarchy is tightly designed. Although that is an orthodox solution, it makes the foremost feature of this machine's high efficiency.

am i saying the way they designed this console put it in same ball park as ps4 no but its VERY efficient and because of that.... it can "punch WAY above its weight."

Of course its efficient, a la the memory hierarchy I was arguing more against the introduction of new FF gear that people where previously suggesting.
 
GC data rate was 2 MB/s to 3.1 MB/s depending on the area of the DVD it was reading.

That's not really indicative here though:

u1kBCUU.gif


Game might look the same, but they changed pretty much all the texture assets for heftier ones.


The game is out now in US on digital form, if you want the blu-ray version you'll have to wait for October 4th still. So people should be testing it installed to the flash ROM or external hard drive (making GC's drive data rate inconsequent); now, it could be a specially lousy drive being used for streaming in this case, or the game really has consistent slowdowns; installs are still a best case scenario for Wii games after all and people that reviewed it reviewed the digital version... same case for people that are playing through it now.

That's the first time I'm reading about it though, and I know people that are playing through the game now; I'll have to ask them.
 
GC data rate was 2 MB/s to 3.1 MB/s depending on the area of the DVD it was reading.

That's not really indicative here though:

Game might look the same, but they changed pretty much all the texture assets for heftier ones.


The game is out now in US on digital form, if you want the blu-ray version you'll have to wait for October 4th still. So people should be testing it installed to the flash ROM or external hard drive (making GC's drive data rate inconsequent); now, it could be a specially lousy drive being used for streaming in this case, or the game really has consistent slowdowns; installs are still a best case scenario for Wii games after all and people that reviewed it reviewed the digital version... same case for people that are playing through it now.

That's the first time I'm reading about it though, and I know people that are playing through the game now; I'll have to ask them.
In the OT people talk about slowdowns during intense naval battles, which occurred in the original verison too iirc
 
Is it running at 1080 p "60 fps" with some drops?

30 fps because Nintendo keyframed every animation. Putting it up to 60 fps would just make the animations cycle through faster. It'd look really weird. Getting around it would require redoing the animations. No simple fix.

I can't say one way or another, but I hope Stevie is right that it's that frame delay effect.
 
30 fps because Nintendo keyframed every animation. Putting it up to 60 fps would just make the animations cycle through faster. It'd look really weird. Getting around it would require redoing the animations. No simple fix.

I can't say one way or another, but I hope Stevie is right that it's that frame delay effect.
Yeah. Seems to be hard to fix. Even Ocarina 3D suffers some animation bugs despite the effort in adding more animation frames. Dodongos spin like crazy for instance.
 
30 fps because Nintendo keyframed every animation. Putting it up to 60 fps would just make the animations cycle through faster. It'd look really weird. Getting around it would require redoing the animations. No simple fix.

I can't say one way or another, but I hope Stevie is right that it's that frame delay effect.

They could of added keyframe interpolation to the engine (something that just about all modern engines have).
 
They could of added keyframe interpolation to the engine (something that just about all modern engines have).
Yes they could have, but there's still a chance of borking it along the way. So they decided not to touch them. Can't blame them. Animation tends to be the most complicated portion of model and asset production. Especially doing it old school.
 
Yes they could have, but there's still a chance of borking it along the way. So they decided not to touch them. Can't blame them. Animation tends to be the most complicated portion of model and asset production. Especially doing it old school.

I highly doubt it's the specifically the animations that are the problem. It's much more likely that all the game logic is hard-coded to 30fps, and all the update and draw code is all muddled together (as was the style of the time).
 
It would be nice to know when and how bad the framerate drops are. It could shed some lights on the limitations of the hardware, given that this game is a 1st party Nintendo product and just an HD version of an old game.
 
Did Halo Anniversary fix the animations? If so, how?
Halo was a PC game from the ground up (well, a mac game, originally, but point stands) those "optimizations probably don't apply as said game didn't have locked framerate, the original build before going console exclusive would have to accept various target specs, existent and to be released after all.

I'm not so sure most Nintendo GC games were coded like that though; as the Wind Waker engine was pretty ground breaking and cutting edge back then. But I know even some Xbox games suffered due to it, some games would go turbo when people hacked 1.4 GHz Pentium 3 Tualatins in.
 
They could of added keyframe interpolation to the engine (something that just about all modern engines have).

how does that work? does it simply add frames with coordinates averaged from the previous and next frame for all data points, or is there some clever way of doing it that accounts for inverse kinematics, center of mass etc?
 
yeah the Iwata ask said it was done in a year. im enjoying the game and i havent noticed them... though im only about 5 hours in. journlist and reviewers try and make too much out of smaller things... lets not make these small framerate issues something to do withthe hardware not being capable or something.

People will use it as an indicator that the GPU and CPU are crap.
They look for anything to feed their point of view.
 
Is eDRAM clock based on Renesas numbers or is the clock based on assumption. Is there any reason why it wouldn't be clocked higher than Latte?

Both. The Renesas parts I see are clocked at 500MHz, plus running at the GPU clock would make the most sense (being on the same clock domain physically would help keep power low, and would mean no missed cycles or bubbles [as an aside that's why Nintendo has liked running everything in proportion, before now]) so ~529, with a side possibility of running at the main memory clock of 800MHz. The two net you 70.4 or 102.4GB/s. But the eDRAM doesn't seem to be connected to the main RAM in the way the 360s was, so I'm thinking GPU clock is more likely.
 
People will use it as an indicator that the GPU and CPU are crap.
They look for anything to feed their point of view.

Or some people will look at when and under what circumstances it happens and make some reasonable deductions.

Not everyone reading this thread is fighting somekind of hearts and minds war.
 
Somewhere in this thread we did a number swap, whoops, lol. 103, not 130, and that was based on the assumption of the eDRAM being clocked higher than the GPU despite being in the same die.
It's 131GB/s if it's 2048bit at 550MHz. The 70 and 103GB/s figures are based on randomly made up clock frequencies. If the eDRAM runs at 550MHz, it can't be 70GB/s. It's either 65.5 or 131.1. If we assume that it needs to be at least 70 for the ROPs, and considering MEM1 isn't just a framebuffer, it's probably 131.

Unless the embedded memory is indeed running at some random clock frequency, then it could be anything.
 
really dont think WW is going to be a good indicator of bad performance, especially if the original contained it. read the Iwata asks and you'll see it was a tiny tiny team who automated much of the conversion.
 
It's 131GB/s if it's 2048bit at 550MHz. The 70 and 103GB/s figures are based on randomly made up clock frequencies. If the eDRAM runs at 550MHz, it can't be 70GB/s. It's either 65.5 or 131.1. If we assume that it needs to be at least 70 for the ROPs, and considering MEM1 isn't just a framebuffer, it's probably 131.

Unless the embedded memory is indeed running at some random clock frequency, then it could be anything.

thank you based wsippel, I knew I wasn't crazy thinking 130ish was a good probability
 
I don't know about pinouts, but the PS2 Graphics Synthetizer eDRAM had 2560 bit width quite a few years ago.

It's possible, it's also expensive.

EDIT: Regarding pinouts, I reckon the pinouts we could determine/ballpark were the DDR3 ones, since they were off-the-shelf parts; for the GPU embedded memory it's most certainly trickier.
 
Isn't the frame rate drops when you're battling the enemies intentional? I could sworn it was said that the 30 frames went with the over all design of the game as well when it was being developed on GC.
 
Isn't the frame rate drops when you're battling the enemies intentional? I could sworn it was said that the 30 frames went with the over all design of the game as well when it was being developed on GC.

Yes, but we don't know if that's what the video is even talking about.
 
Isn't the frame rate drops when you're battling the enemies intentional?
The so-called pause-effect.

I remember people complaining about constant slowdown on the original Wind Waker; those were it. I don't remember sea battles taking a big toll on the framerate, but it was a long time ago.
 
Also, there are three more considerations as to why 131GB/s seems like a good guess:

  • The number of banks on the macros is, in fact, 256 as far as I can tell.
  • Why would they use eight 128bit macros when they could have used four 256Kw x 256b macros?
  • The whole point of pseudo-static RAM is perfect single cycle latency. Running the RAM at any clock other than 550MHz would cancel that benefit out.
 
It's 131GB/s if it's 2048bit at 550MHz. The 70 and 103GB/s figures are based on randomly made up clock frequencies. If the eDRAM runs at 550MHz, it can't be 70GB/s. It's either 65.5 or 131.1. If we assume that it needs to be at least 70 for the ROPs, and considering MEM1 isn't just a framebuffer, it's probably 131.

Unless the embedded memory is indeed running at some random clock frequency, then it could be anything.

So why not go for 4096?
 
Also, there are three more considerations as to why 131GB/s seems like a good guess:

  • The number of banks on the macros is, in fact, 256 as far as I can tell.
  • Why would they use eight 128bit macros when they could have used four 256Kw x 256b macros?
  • The whole point of pseudo-static RAM is perfect single cycle latency. Running the RAM at any clock other than 550MHz would cancel that benefit out.

that does make a lot of sense
 
I haven't seen any framedrops in WW, I bet they're confusing it with the intentional effects when you hit something. Granted I disn't notice the drops in W101, but one thing is a drop from 60 another is from 30. Didn't get to any intense ship battle though.
 
I haven't seen any framedrops in WW, I bet they're confusing it with the intentional effects when you hit something. Granted I disn't notice the drops in W101, but one thing is a drop from 60 another is from 30. Didn't get to any intense ship battle though.

I'm thinking the same. No other review I read mentioned drops. They are confused.
 
It's 131GB/s if it's 2048bit at 550MHz. The 70 and 103GB/s figures are based on randomly made up clock frequencies. If the eDRAM runs at 550MHz, it can't be 70GB/s. It's either 65.5 or 131.1. If we assume that it needs to be at least 70 for the ROPs, and considering MEM1 isn't just a framebuffer, it's probably 131.

Unless the embedded memory is indeed running at some random clock frequency, then it could be anything.

Not random. But obviously not using the same calculation for bandwidth as you are using. I seem to remember blu and yourself talking about that. How are you getting those numbers again?

Also, there are three more considerations as to why 131GB/s seems like a good guess:

  • The number of banks on the macros is, in fact, 256 as far as I can tell.
  • Why would they use eight 128bit macros when they could have used four 256Kw x 256b macros?
  • The whole point of pseudo-static RAM is perfect single cycle latency. Running the RAM at any clock other than 550MHz would cancel that benefit out.

If you add up the columns in a macro, you get 256, yes. The thing that throws me off and prevents me from saying one way or another is that they are in a "sandwich" pattern. The bottom of the columns (the lighter portions) being on either side of what must be some type of path out to the interface on the left of the modules. If we just count horizontal and assume that each "sandwich" is sharing the same path, it would be 128. I'm really struggling to describe my thoughts here. lol

Something I've been thinking about. Do we know if the GPU downclocks in Wii mode? I know the CPU def does, but I don't think we've gotten word on Latte. In trying to figure out the bandwidth of the eDRAM pools, this info might help. If the eDRAM is playing the part of Wii's 24MB 1t-SRAM, then it would have to be on a different clock than the rest of the GPU...if they are doing BC that way.
 
Halo was a PC game from the ground up (well, a mac game, originally, but point stands) those "optimizations probably don't apply as said game didn't have locked framerate, the original build before going console exclusive would have to accept various target specs, existent and to be released after all.

I'm not so sure most Nintendo GC games were coded like that though; as the Wind Waker engine was pretty ground breaking and cutting edge back then. But I know even some Xbox games suffered due to it, some games would go turbo when people hacked 1.4 GHz Pentium 3 Tualatins in.
Halo could unlock its frame rate, yes. But model animations were locked to 30. It's rather jarring to play because of that.
 
Did Halo Anniversary fix the animations? If so, how?
You mean in terms of keyframe interpolation? Nope, because Halo Anniversary is locked at the same 30fps that the original game is, despite CEA being a port of the 60fps-capable Halo PC. It "doesn't need the fix", basically.

If indeed 131GB/s how does it stack up with current/next gen.

1080p games possible or still best suited for 720p+effects?
There's a lot more to choosing a resolution than output bandwidth.

Here's a consideration: the original Xbox is generally seen as rather fillrate limited, and the PS3's RSX's total bandwidth is only about 2/3rds the 360's bandwidth to eDRAM ROPs alone.
Yet the original Xbox was the console of its gen that was pushing quite a lot of 720p games, and IIRC the PS3 has more 1080p titles than the 360 does. In both of those cases, a major deciding factor would have been the lack of small eDRAM pools. The WiiU's 32MB pool might make it an eDRAM-based design that alleviates that issue to an extent, of course. But it's still up to the devs to decide whether it's worth spreading the pixel shading power (and other things) that thin. The render output scheme does look like a compromise between the extreme strengths of the 360's and PS3's design, and so I would guess that in that respect 1080p might look less costly on WiiU. But that's just one part of the pipeline.
 
Status
Not open for further replies.
Top Bottom