Schnozberry
Member
There isn't enough SRAM for 256 ALUs, unless AMD has customized the registers, which seems to be a serious stretch.
WiiU "Latte" GPU Die Photo - GPU Feature Set And Power Analysis
- Thread starter Thraktor
- Start date
- Status
There isn't enough SRAM for 256 ALUs, unless AMD has customized the registers, which seems to be a serious stretch.
frankie_baby
Member
it also has half a gig of gddr5 built in
frankie_baby
Member
random curveball here and i dont know what i'm talking about, could it be possible that the 1mb module of sram (that we know nothing of its purpose outside wii mode) could be used for registers?
Frankfurter
Member
I'm not sure if that's as clear as you make it out to be. For example, the Mobility Radeon HD 5650, based on the same RV830 chip as HD5550, but using all hardware units (400:20:8) has a TDP of 15/19W at 450/650mhz respectively. Which actually sounds like it's within the limits of the Wii U GPU. Again, that's 400 SPUs, i.e. more than we are seeing as the very best case for Wii U.
Schnozberry
Member
Perhaps? I can't answer that.
http://www.anandtech.com/show/4061/amds-radeon-hd-6970-radeon-hd-6950/4
Anandtech said:
VLIW5 gives only ~70% efficiency on average, in a console setting it would achieve 80-81% maybe.
While your quote is probably ridiculously outdated, 50% of 240 shaders is still 120, something at best case would just barely be edged out by Wii U's 128 shaders @ 80% efficiency now they get a small 10% bump thanks to a higher clock, but this means that NFSMWU is pulling out all the stops and reaching it's limit in the launch window just to edge out Xenos' 50% efficiency. If you could pull a quote that is recent that points to Xenos only using 120 shaders on average, that would be something at least, but until I see that Xenos coded to the metal was far inferior to R700. There is just no way Wii U could keep up with a Xenos at a higher efficiency, certainly not reach a higher performance level. If it's VLIW with 160 shaders, it is a very custom design meant to push close to 100% efficiency, of that much I am certain.
The reason the Mobile parts are able to get this low is the parts are bin. So the top 10% or whatever number of chips can hit this power usage. The other chips are sent to desktop or lower clock parts.
This is why mobile chips have such low power usage. So to make this work for a mass product product you would have to throw away most of the chips you made.
Not out dated they were from the xbox summit last year. The docs leaked online. Maybe it wasnt that. I try searching but didnt find anything so it maybe wrong.
http://www.neogaf.com/forum/showpost.php?p=57639606&postcount=4736 Or the professionals in this post know more about this than you or I, and have come to the conclusion that those registers could be more dense than others on the die which often happens with hand layouts according to them.
They might be 6kb or even 8kb, what the people who actually know what the hell they are talking about has said as much, so this point is completely suspect.
Schnozberry
Member
That chip is 146mm at 55nm. If you shrink it to, 40nm do you have enough room for the EDRAM? I'm not sure how to do the math. The Wii U GPU is 146mm with all the added system logic and EDRAM.
Schnozberry
Member
Fair points. I missed that post somehow.
If I make a bunch of vague predictions about Nintendo, you can bet some of them can be interpreted as right in hindsight.
It's ok this thread is moving fast again and that was posted this morning so it's just an interesting point and shows that BG's assessment of the GPU is just as grounded as anyone elses (256 ALUs) of course that means they aren't VLIW5, but it could be custom or 30 ALUs per SPU, though that seems like an odd numbering, it would give Wii U 240 ALUs and allow their more efficient design to directly compete with 8 year old hardware.
The only thing that really goes against it flat out is why would they move from VLIW5 yet stay at dx10.1. Makes no sense....
frankie_baby
Member
well we have had some people saying they go beyond 10.1
Schnozberry
Member
Do we know enough about the API to say what features are missing from being DX11 equivalent?
Only USC-fan knows these things. btw USC-fan did you find that efficiency quote? as you've said maybe you are wrong, but if you are right, it could mean that Wii U is 160 ALUs even VLIW5 and that would be fine. it would still outperform Xenos handily IF that is the case.
BTW VLIW5 is also DX11, it is about the shaders used and we know that it has features beyond DX10.1 so who knows if this is even a knock against it. (hope you didn't mind the jab usc-fan, I just had to point out that your point is factless)
Fourth Storm
Member
So basically you are saying Xenos' configuration is a better setup than VLIW5 even though VLIW5 came after and is obviously more advanced? What makes you think that Xenos has it topped in efficiency when its shader architecture is very similar yet less advanced? You are just confusing people by throwing all these shader numbers and percentages around in your posts. It is clearly apparent that you have a pretty cursory understanding of these things as you are speaking in ways that make broad generalizations and oversimplify what is an extremely complicated topic in hardware engineering and game engine programming. That's ok, but don't speak as if you are any sort of authority on the matter. I can almost guarantee I've read most of the same articles you have concerning VLIW5/VLIW4/GCN, but I acknowledge the limits of my expertise. You ought to try doing the same.
You also really believe that AMD employed an unheard of architecture, that you designed over a known design that developers are familiar with? The whole thing reeks of grasping at straws, I dare say because we were getting closer to the truth and you didn't like the sound of it. It just couldn't be true, so now we are latching on to anything that is remotely possible and claiming it more likely. The same thing happened when arkam came around and was subsequently run out of town, yet he was ultimately vindicated.
Fourth Storm
Member
This is correct, folks. And AMD's embedded MCMs are based on the mobile chips. Probably the same chips just hooked up to an MCM rather than a mobile card.
Building a strawman argument is misrepresenting somebody else's in order to counter it with one of your own. That's exactly what has been going on. I never said that either Latte or Xenos was being maxed out by NFS. Anyone who claims that any game is using this % of the shaders or that % of the shaders is talking out of their ass. It's far more complicated than that, and nobody is qualified to say anything on the matter except the devs. What I did say was that Criterion could probably get better efficiency out of Latte than they did Xenos. This is due to the more modern shaders, higher clock, more RAM, and fast eDRAM.
7 and a half years of experence with Xenos + coding to the metal + the architecture doesn't have THAT limitation.
I don't really care what the reality is of Wii U, I've tried to grow up from the whole Fan bashing bs that is passed around on message boards, my reasoning for Wii U not having 160 shaders is because that many shaders with VLIW5's set up just doesn't have enough efficiency to overcome 50% more shaders Xenos has, certainly not to a level higher than Xenos that we hear often from developers.
What is your point in even attacking me? debates that turn this way are often because one side is loosing ground and feels that they need to slander the other in order to make any statements from that person discredited.
As to the points I've mentioned 180 ALU efficiency that Xenos would have to have in order for 160 ALUs for Wii U to make since. if given 4 or 5 years with the architecture I'm sure VLIW5 could maybe get to 90% efficiency meaning that they are constantly hitting 4.5 slots instead of the 3.4 they did on PC, but that is unlikely to be found out of the box and would still leave Wii U at ~144 ALUs on average, playing catch up to xenos hardly makes sense with what we have seen and known about the poor development tools and unfinished hardware. Nothing I am saying is really that deep or technical it's all surface explanations more or less and I'm a security guard for a living not an engineer, but I didn't just learn about VLIW yesterday, I've been studying it since 2008.
The problem with this is the amount of efficiency they would have to achieve out of the box with no prior experience with the hardware, poor tools and unfinished SDKs. Compared to their 7+ years of experience with Xenos and the obvious limitations of VLIW5. It's not so much that Xenos architecture is better, its that it has more to work with. we aren't saying Xenos has 160 ALUs and Wii U has 160 ALUs, we are saying Xenos is a V6 and Wii U is a V4, there is limitations found in that alone, not to mention that Xenos is deeply understood.
Schnozberry
Member
Does 1 Xenos Shader = 1 Latte Shader in terms of performance? I would find that hard to believe. It's not impossible that their are more ALU's than meet the eye, though.
What the previous poster said about density and hand layouts makes sense. We are operating on incomplete information and trying to deduce from there. No doubt we could be making calculation errors that make our other math inaccurate.
im sorry to call you on this, but you are equally as adamant as he is. this is a discussion, hes allowed (and you) to throw theories around and debate them - why you getting so defensive all of a sudden? you haven't really written anything new or debated, just belittled, but used to contribute loads.
bolded above are really pointless sentences that dont contribute much, or go faaaaaar out of there way to the point.
edit: any comments on this PM me, since dont wanna take up any more space of this thread
Fourth Storm
Member
First of all, very few games are "coded to the metal" these days. Everything goes through an API, whether it be on console or PC. PCs may have more somewhat more overhead but that's pretty much the extent of the difference in most cases (devs like Naughty Dog and Factor 5 have gone a bit further). And you're still doing it by saying these things are however much "shader efficient." Why did you not acknowledge what I wrote regarding PC games being programmed to work on a variety of cards from both AMD and NVidia (with Nvidia being the dominant force in that space, it bears noting)? Of course PC cards are going to end up being less efficient. But you are basically saying developers aren't taking the proposed VLIW5 of Latte into account when programming? "We are only using 128 shaders at a time, but whatever." Similarly, where are you getting the idea that Xenos wouldn't have the same limitation as a VLIW5 part if both were given equal effort from developers?
I've tried to stay away from personal attacks, but you have practically hijacked the thread with a bunch of pseudoscience, and it's a bit aggravating because the level of discussion has taken a nosedive. It's sad that there are probably some people out there that can probably help clarify things further (Marcan has provided alot of information for us that I never would have guessed) but they have either moved on or never had any interest in the first place. What that says about us, well, I won't go there.
I actually have more tidbits that might aid our analysis, but the waters have been so muddied at this point, I wonder if it's even worth it.
Which is why I never try to be exact with this, too many unknowns to speak with absolutes. even my ALU counts are not set in stone, in fact I've pointed to 160, 240, 256 and 320 all as possibilities. Some posters here are more sure about it than they should be and the reality is they are basing their conclusions on as much assumptions as the rest of us and while I am grateful for fourth storm's analyst of the die. He isn't the authority on this and has said himself that he doesn't understand VLIW enough to compare it to Xenos properly, BGassassin disagrees with him, stating the minimum ALU count is 256. My take is that it could be lower but would need to be custom, this isn't some fan boy conclusion, if it was I would never "allow" 160 ALUs to even be suggested, just as some refuse to allow 320 ALUs to be.
So no, 1 Latte shader should not = 1 Xenos shader, but 1 Latte shader has = 1.5 Xenos shaders to even put it on the same level, except out of the box Latte won't be using it's 5th shader often it was purposefully left aside in VLIW5 to be used as a FPU and is a very wide architecture especially when xenos and PCs typically only use 4 or less shaders in a given wavefront (thus VLIW4 was created) this means Xenos is better for handling Xenos code than Wii U would be at it, limiting Wii U's architecture often to only 128 shaders, making the comparison even worse, of course both architectures have idling shaders, but Xenos understanding and being the lead platform would likely make it as efficient or better than GPU7.
ForeverZero
Member
So we can categorically rule out 640 ALU?
Alas..it wasn't meant to be...
Alas..it wasn't meant to be...
Schnozberry
Member
Well, to whatever extent the waters are muddied, new information is always welcome. Your analysis is the best we've seen on this forum, so please don't be discouraged.
chaosblade
Unconfirmed Member
Well, I'm interested, but that doesn't mean much since I don't really have anything to contribute anyway. I just find it all very interesting. Hopefully there are similar breakdowns for the other two upcoming consoles, but given how much is already out there I doubt it.
And it is unfortunate where this topic has gone, I doubt the level of discussion will ever really pick back up.
I think you know what I meant by coding to the metal, it is quite well known that Xenos has far less overhead that PCs thanks to DX11 being used to be compatible with all cards. Your points about developers having put the same effort forward is simply impossible because of unfinished hardware and poor tool sets. as to your point about this thread taking a nosedive, I'd say it's a lot better than a few days ago when people were talking about the possibility of a 3.24GHz clock on espresso via update.
What you know, you don't know, you assume, we have a few facts and a ton of assumptions, what makes your analyst of the Wii U die more factual than BG's? What makes your assumption of performance from 160 ALUs much more of an expert opinion than BG's 256 ALUs? You point at me, but I'm not the only one who doesn't think your analyst of the Wii U GPU is a fact. This thread isn't about one person giving the answers, it is about trying to figure out as a group what Wii U's GPU is.
frankie_baby
Member
If you have information to share its always welcome, some of the best discussion is disagreement so please don't feel put out just because things aren't harmonious
Chronos24
Member
Not sure if this helps, but looking over wikipedia on the radeon families of GPUs, 4000 series did not support Dolby digital and only LPCM and DTS. 5000 family does. WiiU only puts out LPCM, which is what 4000 family puts out. If this makes any difference or narrows what we're looking at down at all great. If this is already known fact shun me 
Fourth Storm
Member
This thread is getting derailed and the old adage about arguing on the internet keeps popping into my mind. Let's just agree to disagree and one day when these things are finally leaked, one of us will buy the other a virtual beer.
Thanks dude. Btw, I just dug up this old quote from CVG. I know, "anonymous devs" and all but with NDA's under effect, it's really the only way to speak out unless you're like Harada and like to play with fire.
So that's that dude, Marcan, and the leaked features sheet all pointing to a pretty conventional Radeon architecture. There may even be more that I'm forgetting.
Fourth Storm: Right, it isn't like I don't like your analyst for a while I even believed it, which is why I figured it had to be custom considering VLIW5's limitations putting it on similar footing with Xenos limitations.
And it isn't like I haven't brought up any valid points, which is the main reason I am defending my posts, they have merit, I'm not an engineer but I pointed out that 160 shaders isn't 160 shaders all the time, something that I feel is being overlooked and stopping discussion in this thread only to be left with whatever conclusions the last person who tried has come to.
frankie_baby
Member
Not sure if this helps, but looking over wikipedia on the radeon families of GPUs, 4000 series did not support Dolby digital and only LPCM and DTS. 5000 family does. WiiU only puts out LPCM, which is what 4000 family puts out. If this makes any difference or narrows what we're looking at down at all great. If this is already known fact shun me
Nintendo use a different audio hardware in there not amd's
You say that like Nintendo made the chip in their own chip factory. Whatever parts that are in it and wherever they are from, AMD and Renesas put them there. So you can't really say its not AMD's since they are the chips manufactoring.
Do we know exactly who manufactured the DSP?
It's not off topic, we are discussing Wii U's GPU and the possibility of the architecture and ALU count. I don't mind disagreeing with you, it's not that I think you are wrong out right, just that your conclusion might be wrong.
As for that quote, yeah what I'm suggesting would still be VLIW and thus conventional, the only real difference is that stuff wouldn't be hanging to be scheduled and the cycle latency would drop drastically to more match Wii U's set up. r700 has a cycle latency that is ridiculously high, Cayman which was much improved is 44 cycles and is still extremely nasty to work with compared to GCN, largely because not enough schedulers closer to the hardware.
Actually iirc the manufacture of the DSP is known they wrote a document and is the reason we know it is clocked at 121.5 mhz (again iirc)
frankie_baby
Member
This thd is getting derailed and the old adage about arguing on the internet keeps popping into my mind. Let's just agree to disagree and one day when these things are finally leaked, one of us will buy the other a virtual beer.
Thanks dude. Btw, I just dug up this old quote from CVG. I know, "anonymous devs" and all but with NDA's under effect, it's really the only way to speak out unless you're like Harada and like to play with fire.
So that's that dude, Marcan, and the leaked features sheet all pointing to a pretty conventional Radeon architecture. There may even be more that I'm forgetting.
I remember that cvg anonymous devs quote, feel free to correct me but isnt AI generally handled with general purpose code not floating point thus espresso should outperform xennon at that so making that quote a bit fishy
Fourth Storm
Member
http://www.techpowerup.com/reviews/Sapphire/HD_6450_Passive/9.html
This was brought up in the beyond3D thread by function. I initially rejected the comparison but now I believe it to be somewhat valid. The green bar represents a card with a 160:8:4 core layout. It runs at 625 Mhz (a bit higher than Wii U) but it is also pushing 44.1 fps in the lowest configuration (1024x768). Mind you, Wii U is running BLOPSII with around that same fps average but at an 880x720 resolution. Now, of course this is not a rock solid comparison since it's CoD4, and while you could say that BLOPSII pushes the engine more, the point could also be made that they've refined their engine over time to handle it. In fact, the 360 hits 60fps for respectable amounts of time in both games. Also keep in mind that the card in question has only a gig of DDR3 at 10.67 GB/s, no eDRAM, and less ROPs than proposed for Latte.
This was brought up in the beyond3D thread by function. I initially rejected the comparison but now I believe it to be somewhat valid. The green bar represents a card with a 160:8:4 core layout. It runs at 625 Mhz (a bit higher than Wii U) but it is also pushing 44.1 fps in the lowest configuration (1024x768). Mind you, Wii U is running BLOPSII with around that same fps average but at an 880x720 resolution. Now, of course this is not a rock solid comparison since it's CoD4, and while you could say that BLOPSII pushes the engine more, the point could also be made that they've refined their engine over time to handle it. In fact, the 360 hits 60fps for respectable amounts of time in both games. Also keep in mind that the card in question has only a gig of DDR3 at 10.67 GB/s, no eDRAM, and less ROPs than proposed for Latte.
Schnozberry
Member
Probably on early kits too, which were hot garbage. I'm sure it's improved since the quote.
Schnozberry
Member
Plus it has the disadvantages of not being on die with the processor, and going through the presumably less efficient directx api.
Fourth Storm
Member
Correct. Now, allow me to draw your attention to the llano die shot in the OP. One huge reason why I think there are 160 shaders is because in my analysis there are 8 TMUs, and a 320:8:8 configuration would be lopsided and ridiculous.
http://images.anandtech.com/reviews/cpu/amd/llano/review/desktop/49142A_LlanoDie_StraightBlack.jpg
Notice the rectangular blocks in the top middle. Those are the L1 texture caches. Now look at the T blocks and S blocks on Latte. Cache is comprised of large pools of SRAM to store the data and other tiny pools of SRAM to function as tags. The left of the L1 caches on Llano looks remarkably like what we see in the S blocks of Latte - a pool of 8 kB and one tag to go along with it. The small SRAM blocks on the right side of the Llano caches seem to have been moved into the T blocks (texture units) on Latte. Both have exactly 16. Coincidence or smoking gun? I'm thinking they probably have something to do with data transfer between TMU and L1 cache; thus they appear in slightly different locations in the slightly different architectures.
It's an apple to oranges comparison, other games are listed in that review, cod4 is cherry picked to show a comparison to black ops 2 its the most unscientific thing I've seen, and I saw it first on beyond 3D, we could compare crysis 2 from there or facry 2 and find that it doesn't handle those games as well as xenos, but instead of a more similar comparison which is still matched up with a faster CPU and a higher clock speed, we are given cod4 a game that was released in 2007? With one that was released 6 months ago and expected to assume it isn't taxing the system differently. Check benchmarks for cod4 through black ops and you'll find performance does get worse.
The Abominable Snowman
Member
Is BG the guy with no tech knowledge but has info and analysis from his insider friend?
Absolutely, not disputing that. I should have made it clearer that I was not disagreeing with the identification of functional blocks based on comparison with other images. Similarly I'm not in a position to disagree with your reasoning on 160 based on the results we've seen from various games. I am making no attempt to argue one way or another as to how many shaders there are or the relative efficiency of modern GPU architectures. Any knowledge I once had in this domain is long gone and dates from a time when men were men and men wrote their own shaders using nothing but a lump of coal and a stick. Or ARB assembly as it was also known.
I agree a combination of the parameter (ALU count, power draw etc) ranges and real world results is as good as we can get until someone leaks. Thought I'd try and help out narrow the parameter range side of the equation by diverting the energy of globo-mega-corp employees from world crushing to the great Wii U hardware debate! Unfortunately they weren't much use at the narrowing bit...
On power consumption, this is again influenced by the layout but can vary dramatically within the same node size depending on the fab process. The Wii U also has the advantage of working to a fixed (relatively low) clock. Most fabs have low power and high performance processes for a given lambda. Do we know the fab and core voltage of the GPU?
Yeah that turn not to be so wrong he quit posting....
TO be fair I was wayyy off too... From sept 14, 2012
http://67.227.255.239/forum/showpost.php?p=42128123&postcount=2747
Jan of this year.
http://67.227.255.239/forum/showpost.php?p=46777838&postcount=1746
Wasnt too far off. Still think we are looking at 160ALU chip or something in between. Still is funny thinking about how I got flamed so hard in these threads. Too funny now...
Fourth Storm
Member
Here's a quote from Marcan's last post:
http://fail0verflow.com/blog/2013/espresso.html
Hmm, so Latte has multiple video output heads on it (presumably 2 - one for TV and one for Gamepad). Where could they be? How about the Q blocks? Odd location perhaps, but they look remarkably like the two small and identical blocks on the very left of the LLano die, which coincidentally also supports 2 video outs (unlike it's successors, which feature true Eyefinity support).
Marcan said:
http://fail0verflow.com/blog/2013/espresso.html
Hmm, so Latte has multiple video output heads on it (presumably 2 - one for TV and one for Gamepad). Where could they be? How about the Q blocks? Odd location perhaps, but they look remarkably like the two small and identical blocks on the very left of the LLano die, which coincidentally also supports 2 video outs (unlike it's successors, which feature true Eyefinity support).
frankie_baby
Member
shouldn't there be 3 though seeing as 2 gamepads are supported?
The Abominable Snowman
Member
He came back but admitted he had no engineering or tech background to speak of and how he was just reiterating stuff from his friend who may or may not have an engineering background themselves.
Which is why I'm confused as to why, after he's been factually wrong so many times, he's still being held in high regard. Is it because he writes huge posts, and too often people just believe what little point he bolded (Regardless of the unlikelihood of the rest of the material)?
I stopped going to these threads, mostly because I was chased out by witch hunters who refused to be realistic in their expectations. Fourth Storm is pretty realistic, and pretty much started these 'investigations' into the hardware, so it's almost hilarious that some of the rambling users have snapped back and begun to reject him, also.
Couldn't they halve the framerate and send alternating frames to each controller?
Is it confirmed that the Wi Fi unit can connect to multiple devices? I'm sure it can if it's very modern, but I forgot where to find the writeup on it.
Doesn't Wii u support 3? 2 gamepads and all...
Fourth Storm
Member
Yeah, but only at 30fps/hz, meaning they've had to jury rig it somehow.
frankie_baby
Member
i'd assume thats due to streaming limitations, will surely need separate video outputs
- Status