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

Status
Not open for further replies.
I think part of his point is that Joe Blow isn't even going to see the difference between PS3/360 and the PS4/720. The fact WiiU falls in the middle means they won't see the difference there either.

If you think otherwise, you're overestimating how much the average person knows and what they look for.

I know what he's saying. I think 2-3 years from now, so at the very least 2nd gen games, Joe Smoe will definitely be able to the difference between PS4 and PS3 games. Wii U is much closer to ps3/360 then the next gen cosoles, so they'll end up seeing the difference there too.

People that think Wii U falls directly in the middle of the current gen consoles and the next gen ones are sorely mistaken. Wii U is 2x 360, where as Durango is about 8x and ps4 is 10x. It isn't all about flops either.

I agree Joe Smoe won't be able to tell the difference between Durango and PS4. Except I do think gamers will be able to tell the difference at games running much higher resolutions and higher framerate.
 
I think Digital Foundry provides good analysis many times. There are however, a few glaring articles that are just flat out wrong. The article they put up about the die shot for example was utter crap. They categorically ruled out the Wii U being 'next gen' even tho they were ignoring 40% of the die shot itself.

And they were all like, "Well we can say, beyond the shadow of a doubt, that Wii U games will not much better than the PS360..."

When it comes to the games I feel like more or less they did good stuff. :)
 
I think Digital Foundry provides good analysis many times. There are however, a few glaring articles that are just flat out wrong. The article they put up about the die shot for example was utter crap. They categorically ruled out the Wii U being 'next gen' even tho they were ignoring 40% of the die shot itself.

And they were all like, "Well we can say, beyond the shadow of a doubt, that Wii U games will not much better than the PS360..."

When it comes to the games I feel like more or less they did good stuff. :)

My friend the problem is that the Wii U do not have YET the software to back this up. It is the first time that I am waiting E3 so nintendo will show a big production game running on the Wii U so I can read what analysis the DF guys will write to back their previous claims.

But DF guys DO NOT provide any good analysis only flame baits for console wars. That's the reasson eurogamer bought them.
 
I think Digital Foundry provides good analysis many times. There are however, a few glaring articles that are just flat out wrong. The article they put up about the die shot for example was utter crap. They categorically ruled out the Wii U being 'next gen' even tho they were ignoring 40% of the die shot itself.

And they were all like, "Well we can say, beyond the shadow of a doubt, that Wii U games will not much better than the PS360..."

When it comes to the games I feel like more or less they did good stuff. :)

No one is ignoring anything. Nothing they posted was flat out wrong. 40% of the die is edram. We have no proof of hidden features. If you have some new info I love read about it. Other than new info it just wishful thinking.

Crazy thing is we have people in this thread saying you won't see much different between ps3 to ps4. That is 10x the power.
 
No one is ignoring anything. Nothing they posted was flat out wrong. 40% of the die is edram. We have no proof of hidden features. If you have some new info I love read about it. Other than new info it just wishful thinking.

Crazy thing is we have people in this thread saying you won't see much different between ps3 to ps4. That is 10x the power.

i've just started to see video gaming as more of a constant evolution, rather than one that is done in "generations". the leap in hardware is significant, and surely by the end of the PS4 lifespan we're going to get some marvels, but i think that a lot of that improvement will be due more to improved software design practices rather than the actual hardware.

bolded one of your statements just because i thought it was funny wording
 
WIth all the GPU discoveries and confirmations one thing that has gone unspoken for is the GX2 API used for it. I am most certainly curious if its truly a match for Open GL 3,3 or higher. Also would love to hear of the Wii U support for Compute Shader and if we may see a port of the TressFX tech.
Well Iwata said the Wii U was capable of GPGPU.
 
No one is ignoring anything. Nothing they posted was flat out wrong. 40% of the die is edram. We have no proof of hidden features. If you have some new info I love read about it. Other than new info it just wishful thinking.

Crazy thing is we have people in this thread saying you won't see much different between ps3 to ps4. That is 10x the power.

Oh I agree that hoping for new info right now might be wishful thinking. Any information would just be a welcomed addition now, however, are you also ignoring the large portion of the GPU that DF ignored? They mentioned the EDRAM so that's not the 40% of the GPU I was referring to. The fact of the matter is that NO ONE (regular consumers like us, and even game journalists) knows about the full feature set of the Wii U and what exactly it can do in terms of graphical prowess. Heck, even some devs didn't know about it until just recently.

Don't get me wrong and throw me in the same pot with those people claiming that the Wii U's "secret sauce" will save it graphically, because I know that it's not as powerful as the PS4 by a long shot.

As for the PS4, the people saying that we won't see much of a difference from the PS3 are just giving knee-jerk reactions I would say. Personally, I thought the PS4 games shown looked amazing.

You guys should be wary of badmouthing DF. I got banned for laughing at them trying to count pixels off of a G4 feed at one E3

Not trying to bash DF. Just saying that they themselves can slip up and make mistakes sometimes. Like I said, I think they provide good analysis for the most part. :)
 
No one is ignoring anything. Nothing they posted was flat out wrong. 40% of the die is edram. We have no proof of hidden features. If you have some new info I love read about it. Other than new info it just wishful thinking.

Crazy thing is we have people in this thread saying you won't see much different between ps3 to ps4. That is 10x the power.
We were not able to identify most of the GPU blocks. The eDRAM is pretty much the only part we understand. We know where the shader clusters are, but we don't understand how they work, as their size and the number of register files makes no sense. The things we (and DF) identified as TMUs might be something else, as they look completely different from traditional TMUs - more logic, less register files. We don't even know what process Nintendo used, which would be quite important to guess the transistor count and draw any conclusions based on power consumption.
 
WIth all the GPU discoveries and confirmations one thing that has gone unspoken for is the GX2 API used for it. I am most certainly curious if its truly a match for Open GL 3,3 or higher. Also would love to hear of the Wii U support for Compute Shader and if we may see a port of the TressFX tech.

TressFX is more likely not supported. While AMD has a tenancy to use console as a testing ground (messed up when you think of it that way), I doubt they were working on this tech in 2011. As for Compute shaders, that was confirmed some time ago that Latte was a GPGPU and had full support. The question only remains is if they will be used.
 
I can understand the other posters feelings towards the DF article. They stated so many incomplete guesses like they were absolute fact and promoted their complete opinion("we can now, categorically, finally rule out any next-gen pretensions for the Wii U") as fact. That single line removed all professionalism from that write-up and destroyed a large chunk of their credibility for me ontop of dismissing over 30% of the chip as irrelevant because they didn't know what it is. As it stands, people look to it solely as a means to bash Nintendo's console as verified proof of weakness as opposed to using it to learn what the Wii U GPU can do. It is forwarding and promoting misinformation.

There needs to be a new and far more professional writeup done by someone else. For as much good as Digital Foundry has contributed, their bias is simply to heavy in this matter. This should be addressed and someone need to compile what has been learned, whether conclusive or inconclusiveb and list it properly according to what it actually is, not what they wish it to be.

One thing that is still bothering me though is why are people, so certain that it is based off of the RV7XX and nothing else? Where was this "confirmed"?
 
Hello all!

I have followed this discussion quite a lot and finally got the possibility to participate.

At least I can help with the comparison die shots because I have spent a lot of time collecting and looking for the best possible die shots available in Internet. In that process I've found higher resolution versions of Brazos/Ontario/Zacate and RV770 die shots than those used for comparison here:
High-res Brazos/Ontario/Zacate APU die: http://www.ixbt.com/cpu/images/amd-bobcat/bobcat-die-hq.jpg
Higher-res RV770 (Ati Radeon HD 4870/4850) GPU die: http://www.pcgameshardware.com/screenshots/original/2008/07/RV770_single_Chip.png
The Llano APU die shot used here is the best one I've found, and for those interested in the looks of VLIW4 Radeon GPU here are two nice high-res Trinity APU die shots: http://www.hitech.co.kr/file/bbs/102/102_764_Second_Generation.jpg and http://electronicdesign.com/content/content/73953/73953_fig1.jpg.

These at least allow us to compare the blocks with higher resolution so that the comparison pictures can be updated and hopefully some new information like similar blocks can be found easier. Here is some information about UVD position of Llano and blocks of Trinity and Brazos so that we can know better what the blocks do. Anyway, Latte should have some kind of UVD and display controller in those unknown blocks, so I hope we can identify them with this additional information.

We should also remember that the unknown Latte blocks should also include the ROPs, memory controller and likely also tessellator together with some other GPU parts for which blocks with similar functions exist in these APU and RV770 dies. Then there are those DSP, ARM and possibly some Hollywood logics for compatibility in Latte that shouldn't appear in any of these APUs in addition to possible custom Nintendo logics. Maybe detailed comparison of these die shots might let us identify some of these blocks, or maybe not because blocks with similar functions may have completely different lay-out in different chips with different FAB proccesses and also base building blocks of GPU may be redesigned with some additional custom functions in Latte. At least all blocks in Latte are different unlike in other GPUs, so the complete chip seems to be optimized to minimize die size instead of just copying the same shader and TMU blocks like in other GPUs.

If you're interested in die shots or if these die shots look beautiful, then please check my die shot collection thread at CPU World forum: http://www.cpu-world.com/forum/viewtopic.php?t=19888. There you'll find a lot of links to beautiful die shots including most AMD, Intel and IBM CPUs. I still have a lot to add there including about a dozen nVidia GPU die shots, so I'll continue expanding the collection.

I also want to really say Thank You to Chipworks, NeoGAF and Marcan because all this has given me so many great die shots to my collection. Please PM me here or at CPU World or contact me otherwise if you have some nice die shots for me that I may not yet have in my collection.
 
Idle curiosity: So how likely is it that this GPU can outperform what is in the PS4(or XB3) in terms of lighting and DoF?
 
Idle curiosity: So how likely is it that this GPU can outperform what is in the PS4(or XB3) in terms of lighting and DoF?
Extremely unlikely because they have bigger overhead.

Even if it's "free" or has reduced hit here, it's just a means to do just that, use freebies in order to close the distance. (meaning you'll be wasting more resources to do it elsewhere, but in the end that just free's up the Wii U to try and close the difference, by not wasting as many resources there.

Then again, that was GC's strategy. It didn't win on raw polygons, but it had so little hit texturing that it featured the best framerate per polycount of it's generation by a great deal; EMBM being able to be applied to almost every surface being a byproduct of that (or the fur shading so abundant in some games, along as the very good water effects for it's time).

If it's free it probably also means it's fixed function, so less customizable it is.


It's a nice thing to have, for sure. "if" it's there. Nintendo loves to have predictable performance, so it just might be the case. Now predictable performance being a Nintendo thing can be well executed (GC for 2001, 3DS right now with it's Maestro feature set) or be badly executed or just out of touch (Wii for 2006, DS). I'm including DS on the last one because the 120k polygon's per frame limit was nuts, you just couldn't go higher and if you did instead of an error message you'd start having missing geometry: fun.


Nice to see you over here Birdman. welcome!
 
Idle curiosity: So how likely is it that this GPU can outperform what is in the PS4(or XB3) in terms of lighting and DoF?

...you're joking, right?

If the rumors are to be believed, the PS4 GPU is of the Radeon 8000 series. It'll outclass the Wii-U GPU in every category imaginable.
 
TressFX is more likely not supported. While AMD has a tenancy to use console as a testing ground (messed up when you think of it that way), I doubt they were working on this tech in 2011. As for Compute shaders, that was confirmed some time ago that Latte was a GPGPU and had full support. The question only remains is if they will be used.
TressFX is actually based on the DirectX11 feature set and not some exclusive feature. It's just written by them and optimized for GCN architecture, but it'll run on any DirectX11 card, including nvidia ones, providing it has enough GFlops/Stream processors to support it.

DirectX11 is a 2009 implementation, so that means theoretically Radeon HD 5xxx, GeForce 4xx and over can support it; probably more than theoretically for the flagship cards of that era. The real question for Wii U should be whether those resources are well applied on that effect (or rather, how heavy is the hit of doing it) and if that code is well written enough for VLIW5.

Official support is most likely not in the cards or something they'll advertise, but if it such compute shader is open sourced then support is not needed anyway, just implementation.
 
Hello all!

I have followed this discussion quite a lot and finally got the possibility to participate.

At least I can help with the comparison die shots because I have spent a lot of time collecting and looking for the best possible die shots available in Internet. In that process I've found higher resolution versions of Brazos/Ontario/Zacate and RV770 die shots than those used for comparison here:
High-res Brazos/Ontario/Zacate APU die: http://www.ixbt.com/cpu/images/amd-bobcat/bobcat-die-hq.jpg
Higher-res RV770 (Ati Radeon HD 4870/4850) GPU die: http://www.pcgameshardware.com/screenshots/original/2008/07/RV770_single_Chip.png
The Llano APU die shot used here is the best one I've found, and for those interested in the looks of VLIW4 Radeon GPU here are two nice high-res Trinity APU die shots: http://www.hitech.co.kr/file/bbs/102/102_764_Second_Generation.jpg and http://electronicdesign.com/content/content/73953/73953_fig1.jpg.

These at least allow us to compare the blocks with higher resolution so that the comparison pictures can be updated and hopefully some new information like similar blocks can be found easier. Here is some information about UVD position of Llano and blocks of Trinity and Brazos so that we can know better what the blocks do. Anyway, Latte should have some kind of UVD and display controller in those unknown blocks, so I hope we can identify them with this additional information.

We should also remember that the unknown Latte blocks should also include the ROPs, memory controller and likely also tessellator together with some other GPU parts for which blocks with similar functions exist in these APU and RV770 dies. Then there are those DSP, ARM and possibly some Hollywood logics for compatibility in Latte that shouldn't appear in any of these APUs in addition to possible custom Nintendo logics. Maybe detailed comparison of these die shots might let us identify some of these blocks, or maybe not because blocks with similar functions may have completely different lay-out in different chips with different FAB proccesses and also base building blocks of GPU may be redesigned with some additional custom functions in Latte. At least all blocks in Latte are different unlike in other GPUs, so the complete chip seems to be optimized to minimize die size instead of just copying the same shader and TMU blocks like in other GPUs.

If you're interested in die shots or if these die shots look beautiful, then please check my die shot collection thread at CPU World forum: http://www.cpu-world.com/forum/viewtopic.php?t=19888. There you'll find a lot of links to beautiful die shots including most AMD, Intel and IBM CPUs. I still have a lot to add there including about a dozen nVidia GPU die shots, so I'll continue expanding the collection.

I also want to really say Thank You to Chipworks, NeoGAF and Marcan because all this has given me so many great die shots to my collection. Please PM me here or at CPU World or contact me otherwise if you have some nice die shots for me that I may not yet have in my collection.

Thanks for dropping by, Birdman. I've come across your die shot collection before, it really is a fascinating resource for hardware geeks like myself.

We don't even know what series Latte is part of, though.

To be honest, at this point I don't even think it's worth considering Latte as a member of any particular AMD GPU series. Given the heavy customisation of the chip, I feel it makes more sense to just think of it as a GPU that sits in parallel to the 4000-6000 series. As far as we know, it could be more different to any of them than they are to each other.
 
Extremely unlikely because they have bigger overhead.

Even if it's "free" or has reduced hit here, it's just a means to do just that, use freebies in order to close the distance. (meaning you'll be wasting more resources to do it elsewhere, but in the end that just free's up the Wii U to try and close the difference, by not wasting as many resources there.

Then again, that was GC's strategy. It didn't win on raw polygons, but it had so little hit texturing that it featured the best framerate per polycount of it's generation by a great deal; EMBM being able to be applied to almost every surface being a byproduct of that (or the fur shading so abundant in some games, along as the very good water effects for it's time).

If it's free it probably also means it's fixed function, so less customizable it is.


It's a nice thing to have, for sure. "if" it's there. Nintendo loves to have predictable performance, so it just might be the case. Now predictable performance being a Nintendo thing can be well executed (GC for 2001, 3DS right now with it's Maestro feature set) or be badly executed or just out of touch (Wii for 2006, DS). I'm including DS on the last one because the 120k polygon's per frame limit was nuts, you just couldn't go higher and if you did instead of an error message you'd start having missing geometry: fun.


Nice to see you over here Birdman. welcome!

Didn't the GC prove to be capable of the highest polpygon count that gen? If I remember correctly, it pulled of 20 million at 60 FPS in a real game as opposed to 12 million at 30 FPS in the Xbox1 and 10 million on the PS2. I'm pretty sure it did win in raw polygons.

I feel it makes more sense to just think of it as a GPU that sits in parallel to the 4000-6000 series. As far as we know, it could be more different to any of them than they are to each other.

This right here should be where we stand on the official analysis of Latte then. Not what DF wrote.
 
It is a highly customized GPU, but as has already been discussed the documentation lists it like an RV7xx chip / VLIW
The leaked spec sheet lists the feature set as identical to or a superset of R700, it doesn't really detail the actual architecture.
 
There was a spec sheet sent out to developers before the console even launched.

Than why does this die shot look absolutely nothing like an RV700? Even if it was based on it, with some modification, we should see it.

Even Chipworks said its a completely custom chip. And they analyse these chips for a living.
 
RV8xx is also no more than a linear R7xx revision leaving core structure intact; in past revisions like the Radeon 9500/9700 to the 9600/9800 that would be a small revision of sorts, reflected in the naming changing from R300 to R350 and not a full generation nomenclature. Of course by the time R7xx launched nomenclature didn't work that way anymore, with available cards instead being spanning all over the R7xx numbering logic all up to R790, but point stands.

Regardless, R8xx supports everything newer cards do and it's pretty much the same chip. (well, minus OpenGL 4.3 and other stuff AMD added on the side, some of which Wii U seems to have, like the EyeFinity thing); that's enough to put the R7xx definition in Jeopardy, it's probably closer to an R8xx, perhaps even some R9xx implementations.

Also GCN is at core a R9xx/HD 69xx/VLIW4 evolution.
 
The leaked spec sheet lists the feature set as identical to or a superset of R700, it doesn't really detail the actual architecture.

Than why does this die shot look absolutely nothing like an RV700? Even if it was based on it, with some modification, we should see it.

Even Chipworks said its a completely custom chip. And they analyse these chips for a living.

I was only suggesting that the original base is related to VLIW, which I'd say by its documentation provided by Nintendo is likely (it was a copy+paste job, almost). Not that it isn't a custom chip. Looking at the die shots proves it's not anything near a vanilla R7xx.

As far as detailing the architecture - the pre-launch spec sheet that was leaked didn't even go deep enough to mention anything specific about performance, let alone architecture. lol

Looking at the chip magnified, it's clear it's been modified by the engineers responsible.
 
Didn't the GC prove to be capable of the highest polpygon count that gen? If I remember correctly, it pulled of 20 million at 60 FPS in a real game as opposed to 12 million at 30 FPS in the Xbox1 and 10 million on the PS2. I'm pretty sure it did win in raw polygons.
When it came to RAW untextured polygons both PS2 and Xbox beat it; RAW being a benchmark for untextured polygons, which is useless. Same for RAM bandwidth and other things, GC was very well designed, neck to neck with Xbox if not surpassing it in a lot of areas, but never more powerful on paper. It's weapon was doing things in a very predictable, little hit way.

The thing was a texturing beast though, thanks to good optimization, design, little bottlenecking, compression support and ETB.
 
TressFX is more likely not supported. While AMD has a tenancy to use console as a testing ground (messed up when you think of it that way), I doubt they were working on this tech in 2011. As for Compute shaders, that was confirmed some time ago that Latte was a GPGPU and had full support. The question only remains is if they will be used.

TressFX is a software feature, not hardware. It works on anything with Compute Shader support; it's more than likely perfectly possible to write a version that the Latte GPU could process, the question is whether or not it would be worth the performance hit, as the effect is said to be very demanding on even high end PC parts.
 
Didn't the GC prove to be capable of the highest polpygon count that gen? If I remember correctly, it pulled of 20 million at 60 FPS in a real game as opposed to 12 million at 30 FPS in the Xbox1 and 10 million on the PS2. I'm pretty sure it did win in raw polygons.
Nope.

Gamecube
  • Graphics processing unit:
  • 162 MHz "Flipper" LSI (co-developed by Nintendo and ArtX, acquired by ATI)
  • 180 nm NEC eDRAM-compatible process
  • 8 GFLOPS
  • 4 pixel pipelines with 1 texture unit each[15]
  • TEV "Texture EnVironment" engine (similar to Nvidia's GeForce-class "register combiners")
  • Fixed-function hardware transform and lighting (T&L), 20+ million polygons in-game[18]
  • 648 megapixels/second (162 MHz × 4 pipelines), 648 megatexels/second (648 MP × 1 texture unit) (peak)
  • Peak triangle performance: 20,250,000 32-pixel triangles/s raw and with 1 texture and lighting
    [*]or 337,500 triangles a frame at 60 fps
  • 8 texture layers per pass, texture compression, full scene anti-aliasing[18]
  • 8 simultaneous hardware light sources, up to 32 software light sources
  • Bilinear, trilinear, and anisotropic texture filtering
  • Multi-texturing, bump mapping, reflection mapping, 24-bit z-buffer
  • 24-bit RGB/32-bit RGBA color depth

Xbox
  • GPU and system chipset: 233 MHz "NV2A" ASIC. Co-developed by Microsoft and Nvidia.
  • Geometry engine: 115 million vertices/second, 125 million particles/second (peak)
  • 4 pixel pipelines with 2 texture units each
  • 932 megapixels/second (233 MHz × 4 pipelines), 1,864 megatexels/second (932 MP × 2 texture units) (peak)
  • Peak triangle performance (32pixel divided from filrate): 29,125,000 32-pixel triangles/s raw or w. 2 textures and lit.
  • 485,416 triangles per frame at 60 frame/s
    [*]970,833 triangles per frame at 30 frame/s
  • 8 textures per pass, texture compression, full scene anti-aliasing (NV Quincunx, supersampling, multisampling)
  • Bilinear, trilinear, and anisotropic texture filtering
  • Similar to the GeForce 3 Ti500 PC GPU in performance
 
Nope.

Gamecube
  • Graphics processing unit:
  • 162 MHz "Flipper" LSI (co-developed by Nintendo and ArtX, acquired by ATI)
  • 180 nm NEC eDRAM-compatible process
  • 8 GFLOPS
  • 4 pixel pipelines with 1 texture unit each[15]
  • TEV "Texture EnVironment" engine (similar to Nvidia's GeForce-class "register combiners")
  • Fixed-function hardware transform and lighting (T&L), 20+ million polygons in-game[18]
  • 648 megapixels/second (162 MHz × 4 pipelines), 648 megatexels/second (648 MP × 1 texture unit) (peak)
  • Peak triangle performance: 20,250,000 32-pixel triangles/s raw and with 1 texture and lighting
    [*]or 337,500 triangles a frame at 60 fps
  • 8 texture layers per pass, texture compression, full scene anti-aliasing[18]
  • 8 simultaneous hardware light sources, up to 32 software light sources
  • Bilinear, trilinear, and anisotropic texture filtering
  • Multi-texturing, bump mapping, reflection mapping, 24-bit z-buffer
  • 24-bit RGB/32-bit RGBA color depth

Xbox
  • GPU and system chipset: 233 MHz "NV2A" ASIC. Co-developed by Microsoft and Nvidia.
  • Geometry engine: 115 million vertices/second, 125 million particles/second (peak)
  • 4 pixel pipelines with 2 texture units each
  • 932 megapixels/second (233 MHz × 4 pipelines), 1,864 megatexels/second (932 MP × 2 texture units) (peak)
  • Peak triangle performance (32pixel divided from filrate): 29,125,000 32-pixel triangles/s raw or w. 2 textures and lit.
  • 485,416 triangles per frame at 60 frame/s
    [*]970,833 triangles per frame at 30 frame/s
  • 8 textures per pass, texture compression, full scene anti-aliasing (NV Quincunx, supersampling, multisampling)
  • Bilinear, trilinear, and anisotropic texture filtering
  • Similar to the GeForce 3 Ti500 PC GPU in performance

Aren't those theoretical numbers?
 
Yes both Sony and MS are famous for posting pie in the sky theoretical numbers.

Neither Xbox or Gamecube hit their numbers 100%, but EVEN IF we assume GC hit their numbers 100% because of the fixed functions, it would still be less than Xbox if it was only capable of pushing 60% of those numbers.
 
Status
Not open for further replies.
Top Bottom