Increased memory bandwidth if I'm remembering right.
-
Hey, guest user. Hope you're enjoying NeoGAF! Have you considered registering for an account? Come join us and add your take to the daily discourse.
WiiU "Latte" GPU Die Photo - GPU Feature Set And Power Analysis
- Thread starter Thraktor
- Start date
- Status
Increased memory bandwidth if I'm remembering right.
Toki Tori = Two Tribes
Heavy is about as tech-savvy as... well, any person or thing you can imagine that has no understanding of it in the slightest.
No, it shows that Capcpom already had a tuned engine set up for ps360 (RE5/6) and not for WiiU. I think they wanted to port this game with the smallest possible investment.
ElectricKaibutsu
Member
How do people watch that video and get the impression the Wii U version is worse? The guy playing says the Wii U version is smoother, quicker, sharper and just all around better.
Cosmonaut X
Member
The Two Tribes tweet was:
...which they followed up with:
They were asked whether it was related to texture compression or streaming but said they couldn't say - presumably because discussion of hardware features like that is under NDA.
...which they followed up with:
They were asked whether it was related to texture compression or streaming but said they couldn't say - presumably because discussion of hardware features like that is under NDA.
I hope this analysis has been helpful to some. It is my opinion that from the TMU count alone we can extrapolate that Latte contains 160 shaders. Anything more would be unbalanced and nonsensical. Once we get past the hangup that the SPUs and the TMUs are not in direct contact with one another, it becomes quite easy to see that Wii U features two fairly standard SIMD engines/cores. We must give up any pretense of a dual setup engine, and it should be remembered that the only GPUs to necessitate such a configuration were absolute behemoths with many times the shading capability of even high estimates for Latte.
As for the shaders themselves, I never claimed to have all the answers and I can't name exactly why they are so much larger than would seem to be necessary for 160 40nm SPUs. I've thrown out a few guesses in previous posts and I am sure that there have been a few modifications/shortcuts made especially for the Wii U hardware. Wsippel linked to an interesting article on thread interleaving - could be, but who knows? We could speculate forever (and that might be fun), but there's nothing else we can really say about them without actual Wii U developer insight. At this point, I don't know if I have much else to say about the Wii U GPU. I have attempted to analyze to the best of my abilities, without bias. People can believe or disbelieve as they please, just as I am sure that new game footage at E3 will have various folks claiming that the Wii U is both more and less capable than PS360. Ok, I think I'm going to go play some games now.
Interesting as always, thanks. For all that backlash over just the suggestion of 160 shaders, it seems to me our most reasonable guess now. It's funny, in a sort of way, that the 8th gen consoles will have 8, 16, and 32 ROPs, each doubling the last. I wonder how important that will be. I suppose it only makes sense with relative shader counts, no point in lots of either without both, shaders and ROPs. Hm, the old 2900xt even had 16 ROPs. But on the other hand they haven't scaled up all that much since then, the high end cards having 64 now.
hehe
My common sense tells me it makes no sense for a modern GPU, with a low power draw, taking into consideration heat and older GPU's on the same process, to have MUCH larger SPU's. If someone can give me an explanation as to why this would make sense for Nintendo, then i might buy it. Until then, it's bogus to me. Nintendo doesn't jus do shit like this for no reason. So either the premise is plain wrong, or there has to be a good reason for it.
efyu_lemonardo
May I have a cookie?
I feel like I'm missing the point. Are you implying there's a legitimate reason lots of Wii games looked like they weren't pushing the hardware, or are you implying comparisons to PS2 games are inaccurate and based on superficial observations?
Was PS2 an overall better environment for developers?
Shit if I know, really. But as Fourth said, why have so many more pixel shaders with so few texture mapping unit, or ROPs? Maybe they're bulked up for compute. But a higher count doesn't appear to make sense.
Shin Johnpv
Member
But it also doesn't make sense for the ALU's in the Wii U GPU to be significantly larger. You have two extremes of things not making sense which to me means trying to make any claim on shader count is folly.
They wouldn't just make them larger for the hell of it, it makes very little sense.
Good to see people are still looking at the GPU and trying to figure out what makes it tick. However on this point I don't follow your reasoning, shader units don't have to have any correlation to TMU's, since they can be used for more than simply feeding TMU's.
It certainly could be, but it's not like 130k would be all that outlandish to begin with:Not sure why this is being quoted. Especially when it could easily be a high resolution mesh never used in game.
The model is even being shown as quads whereas video games typically render out triangles (unless you're a Sega Saturn).
123k polygons. And yes, it's the ingame model.
You're making huge assumptions here without any real evidence. For all we know, the game could be CPU limited, which is why there is so more slow down. Even Lego City, which is Wii U exclusive, has slow down and pop up issues according to many reviews.
Doc Holliday
SPOILER: Columbus finds America
Not sure why this is being quoted. Especially when it could easily be a high resolution mesh never used in game.
The model is even being shown as quads whereas video games typically render out triangles (unless you're a Sega Saturn).
Most artists use quads when modeling these days , cleaner than tris. All quads are two triangles so its automatically converted to all tris by the engine. Just gotta make sure the edges are turned correctly. So you'll probably always see quad models when seeing artists screenshots don't say anything about wether its a game asset or not.
megabytecr
Member
This is why I brought up what Two Tribes tweeted. They made it sound like something surprising and a very cool feature. I dont think they are very amateur to tweet a feature like this that exist on every standard GPU out there. Also this is expanded by Ancel's statements about infinite memory.
That is why I thought this could be relevant to the RAM in the Wii U compared to the other consoles and a way Nintendo found to counter for the RAM difference vs PS4/Xbone.
I do programming, just not for games, but I am trying to be logical about it.
thanks for all your work. cheers.
That makes no sense whatsoever. You do realize that the 3DS only has a dual core 266MHz ARM11, with half a core reserved for OS functions? Espresso is a completely different animal.
Its great that people like Forth Storm are spending time comparing Latte to other GPU's to try to figure out exactly what makes it tick. But I do think firm conclusions are being arrived at far to easily and conclusively based on speculation.
As you say the size of the shader blocks are far larger than should be the case for 160 shaders, that's a fact. Now there are theories on why that could be the case, but there needs to be good reason to really believe its 160 shaders to warrant those theories and at the moment there just isn't. There's a suggestion that the GPU may have 8 TMU's which may be too little to make 320 shaders worth while. Well first of all we have to remember that 8 TMU's is still speculation. Second shader units can be used for more than simply feeding TMU's anyway (Nintendo themselves highlighted GPGPU for example).
Again its good to see the speculation on the GPU, but I can't agree with coming to a firm conclusion based on it, especially when there are facts that simply don't fit with that conclusion.
megabytecr
Member
As others have stated, we all appreciate Fourth Storm's analysis, but as he states it is not conclusive as the size of shaders has no explanation as of now. So no, no conclusions must be made imo.
As I suggested, perhaps it bolsters compute in some way. From other GPU architectures we do know what "fat" vs "skinny" shaders can do for compute, look at Fermi vs Kepler, the latter went for smaller simpler and more numerous shaders which helped in gaming, but the compute performance dropped below their own previous generation.
We don't know, because we don't really know what to count. My guess is 2048bit, so 131GB/s for MEM1. MEM0 could be anything. Doesn't help that it's split in two, either. And it's apparently managed by the OS - we don't know what it's used for, so it might not even matter to begin with.
iirc the odds were either 70 or 130, neither confirmed yet.
Meccanical
Member
As an outsider looking in, he seems to be one of only a few people that's trying to determine what the Wii U is rather than dream about what they want it to be. His posts are way more conclusive then half the things I'm seeing in this thread.
With respect to Forth Storm there's nothing conclusive there with regards to shader count or even TMU count, its very interesting speculation and much appreciated by everyone here, but it is speculation.
I also don't see why a perfectly legitimate and significant piece of info (like the size of the shader blocks) should be put down as people dreaming about what they want the GPU to be. I'm bewildered as to why people seem so eager to ignore/explain away a solid fact in favour of speculation that doesn't definitively point to any one conclusion anyway.
I hope Forth Storm doesn't think anyones trying to put down his work, because I don't think they are, I'm perfectly open to the possibility of there being 160, 256 or 320 shaders, any one is possible.
It only "makes no sense" if you assume the 3DS version is running the exact same code.
The console version is different , if it was simply an up-scaled carbon copy version of the 3DS release, it would be running closer to 60hz.
How exactly is it "different"? It's the exact same game with higher resolution maps and textures. It's not like they introduced some high end physics calculations or throw a hundred enemies with vastly improved AI at you all of a sudden. Nothing in the game would explain why a CPU that's at least 10 to 20 times more powerful suddenly runs into problems.
M°°nblade;59551313 said:
Strawman argument. I made no such claims in the way that you state them.
I'm guessing there was a single BETA demonstration out there that showed a hiccup or two at some point, and like with all things Nintendo hardware related, people immediately try to promote the lowest demonstrations as the maximum capability and all perceived problems as the fault of the hardware.
blu
Wants the largest console games publisher to avoid Nintendo's platforms.
Frankly I'm not convinced ; ) I'll elaborate when I get home tonight.Finally, I come back to the subject of the TMUs and L1 texture cache. What I previously identified in the Llano die has once again been confirmed with the Brazos annotation. I have even used my skills at MS Paint to make it crystal clear for you all.
On Brazos, we see the same basic SRAM configuration which I previously used to link the L1s in Llano to the S Blocks in Latte. We have the 2 longer rectangular banks + a much smaller squarish bank (In Brazos, the long banks are strangely staggered, with one slightly longer as well. This is a slight variation and hardly enough to discount the similarities). Right next to these, we have the 16 other small SRAM banks. I have speculated that these might be the cache tags. Whether they are exactly this or not, it is pretty obvious that they are cache related. I bet that if we were to get a more clear picture of the RV770 die, we would see these 16 banks in the TMUs, as they are on Latte. These variations exist across the different architectures. In the Llano photo linked to in the OP, for example, those 16 banks are on the right side of the L1 blocks while the other aforementioned banks are on the left. The bottom line is that this SRAM configuration is indicative of being texture cache and TMU related. The common number of banks points directly to there being 8 TMUs on Latte, just as on Brazos.
BG may be seen as "a dreamer", but Blu, Wsippel among others, are not. I value their input as much or more than that of Fourth Storm. They never ran with 160 SPU as far as i know.
Cosmonaut X
Member
Exactly. If you were going to point the finger at a component to "blame" for the stuttering in RE:R, surely it would be the GPU rather than the CPU?
However, given the same stuttering is found in RE:R and MH3 Ultimate, both have handheld roots, and neither are doing anything terribly demanding, surely it's more likely to be an issue with an unoptimised MT Framework not playing nice with the Wii U hardware?
JordanN
Banned
130k isn't the norm though. And Wii U has alot to prove that it's even 2x better than PS3/360 (or better yet, what does the game look like to consider 130k impressive? It was PS3/360 at one point so what if that's even the same model?).
Who cares about completely baseless speculation like that? The Bayonetta model had far fewer polygons, so there is nothing to even suggest this. I assume next we will be trying to figure out if its beyond xenos as far as feature set right? Or as powerful as vita? 130k polygon model isn't outside the realm of ps3 or 360 but Wii U does allow benefits that those consoles were simply not capable of or had to work much harder to fake.
JordanN
Banned
Do you really believe Platinum has to reuse the same model? Sequels can change/improve assets all the time.
And we have yet to see how far these capabilities can take Wii U.z0m3le said:
Just saying for people not to use the 130k model as being conclusive when we know little about the game.
What else would a 130k model be used for exactly? It's extremely low for a cut scene model, and you also suggested that the game might have a high poly character model at the sacrafice of everything else, but as far as we know, that boss at the end of the trailer was also in game and looked pretty impressive. Just the modern sharers and extra memory would make the Wii U games look better than 360 games ever could, I mean more than double the ram, triple the cache, taking the gpu performance out of the equation and you still come to the same conclusion, Wii U has more to offer the gamer from a visual stand point than last gen consoles.
It's clearly using more detailed 3d character models, improved lighting, AA and completely upgraded shaders. That alone would require rewrites to many parts of the engine. And honeslty it's not the exact code therefore you can't really make 'like for like' comparisons with the 3DS build. You can however compare the PC and console versions because they are all running off the updated build of the game.
I said the game might be CPU limited purely as an example. The main point I was making, was that unless you know exactly what's going on under the hood, you can't be so certain the Wii U build isn't very optimized. We know that Lego City Undercover which is a Wii U only exclusive has frame rate issues. And a large number of the multi-platform games also have more slow down on the Wii U. Also NFS which has updated graphics on Wii U, but supports less players online.
Are we going to start making assumptions that all these people are just shoddy coders too lazy to optimize for Nintendo?
Chronos24
Member
Um actually... Not so blunt as that but factoring in funding, expectation of sales, and size of team (skill of team) then yeah. Developers I feel at least on WiiU dont expect high sales for certain games therefore put less effort in. You get high sales then the next project may be better in every aspect with regards to what I just listed.
The boss at the end of the trailer(that people who were anti-Wii U were insisting was just CG) was confirmed to be an in game monster.
Your argument is filled with fallacies. You are using facts to make claims that do not logically follow.
No aspect of the CPU would cause cars to be dropped form multiplayer. Only a GPU issue would cause that and the GPU is clearly better than the PS3/360's for them to do all of the upgrades they did. You don't draw game models on the CPU. The only thing the CPU is used for is general games code, A.I. and physics. The single player campaign would be more taxing on the CPU than multilayer because it has to calculate the physics over a wider array and the A.I. for the cars. If it were a CPU issue then their would have been far more cuts than a few multiplayer cars. At worst, the CPU would lead to them having the scale back the physics but that didn't happen.
The car reduction was more than likely the result of either network issues or EA simply wanting to shortchange Nintendo some more which far more likely. Sonic All-Stars racing has no such problem. Black Ops on the Wii U has no such problem. Tekken Tag on the Wii U has no such problems.
Your argument is filled with fallacies. You are using facts to make claims that do not logically follow.
No aspect of the CPU would cause cars to be dropped form multiplayer. Only a GPU issue would cause that and the GPU is clearly better than the PS3/360's for them to do all of the upgrades they did. You don't draw game models on the CPU. The only thing the CPU is used for is general games code, A.I. and physics. The single player campaign would be more taxing on the CPU than multilayer because it has to calculate the physics over a wider array and the A.I. for the cars. If it were a CPU issue then their would have been far more cuts than a few multiplayer cars. At worst, the CPU would lead to them having the scale back the physics but that didn't happen.
The car reduction was more than likely the result of either network issues or EA simply wanting to shortchange Nintendo some more which far more likely. Sonic All-Stars racing has no such problem. Black Ops on the Wii U has no such problem. Tekken Tag on the Wii U has no such problems.
JordanN
Banned
You mean the same boss that was on screen for no more than 3 seconds surrounded by pitch black?
A cut scene model doesn't have to be a specific number (especially if the in game model happens to be lower).
I never said Wii U was not better than Xbox 360. I'm only against point blank judgements off what is clearly circumstantial evidence.z0m3le said:
This is not circumstantial. It clear, verified photographic proof.
http://beyond3d.com/showpost.php?p=1699978&postcount=1174
This is also not circumstantial. This was confirmed in game. There are also ran effects and a large object in the background with water sliding effects on it.
Fourth Storm
Member
Krizz, net code and whatnot is run off the cpu. That is what we are talking about, no? Online multiplayer? I don't think the actual issue is rendering more cars.
Looking forward to what blu has to say regarding the TMUs. Just to clarify, except for some of the acronym identifications on Brazos, all thoughts/conclusions were my own. Tried to word my post carefully, but it doesn't hurt to restate that. I know everyone has their own thoughts.
Glad people enjoy the analysis. It's a fun passtime.
Looking forward to what blu has to say regarding the TMUs. Just to clarify, except for some of the acronym identifications on Brazos, all thoughts/conclusions were my own. Tried to word my post carefully, but it doesn't hurt to restate that. I know everyone has their own thoughts.
Glad people enjoy the analysis. It's a fun passtime.
If you are looking for facts then you are in the wrong thread, everything in here is drawn conclusions based on speculation.
That's all GPU stuff, not CPU stuff. If anything, the Wii U port should even free CPU resources, because it enables things like GPU skinning.
That "lazy" shit ruins discussions all the time. The main reasons for inferior ports are the lack of familiarity with the architecture and tools and - most of all - RoI. If the RoI isn't there, no publisher will assign their A team and give them enough time to figure everything out and do a proper job. It's just not going to happen.
frankie_baby
Member
Krizz, net code and whatnot is run off the cpu. That is what we are talking about, no? Online multiplayer? I don't think the actual issue is rendering more cars.
Looking forward to what blu has to say regarding the TMUs. Just to clarify, except for some of the acronym identifications on Brazos, all thoughts/conclusions were my own. Tried to word my post carefully, but it doesn't hurt to restate that. I know everyone has their own thoughts.
Glad people enjoy the analysis. It's a fun passtime.
Is net code general purpose or floating point?
Fourth Storm
Member
Not sure. Maybe someone with dev experience can chime in.
Krizz, net code and whatnot is run off the cpu. That is what we are talking about, no? Online multiplayer? I don't think the actual issue is rendering more cars.
Looking forward to what blu has to say regarding the TMUs. Just to clarify, except for some of the acronym identifications on Brazos, all thoughts/conclusions were my own. Tried to word my post carefully, but it doesn't hurt to restate that. I know everyone has their own thoughts.
Glad people enjoy the analysis. It's a fun passtime.
Most Wanted U must have had some extra sloppy net code then, because those problems are not present any other Wii U online game.
All of the first hand reports I've seen for Black Ops 2 state no issues. The same with Sonic and SEGA All-Stars racing.
I'm not a prof. dev per say, but I am a prof. web developer and i have much experience with net code and networking. You can make it floating point or integer code if you wish. Its your code after all. Though, most of the work is done by the server, not the client. Also, most of the work on the client end is done on the network card/modem, not the CPU or else modems wouldn't be needed.
Unless the game was being hosted, run and routed on a Wii U on top of having unoptimized code, then I don't see how this could be an issue that would lead to it dropping 2 cars from multilayer where no other game has had to do such a thing. Maybe they intentionally networked the game on old, crappy servers? That's the only reasonable explain nation I can see if this were a CPU problem related to net code.
Given E.A.'s history, I would not doubt that it was an arbitrary decision done for no other reason than to disparage Wii U owners, the same as pricing is 60 dollars and then giving $30 discount on their own proprietary Origin network on launch day and the same as announcing that they would no further support the game with DLC before it even touched a shelf.
Schnozberry
Member
Injustice has solid multiplayer too with dropped frames and what not only really occurring when someone has a high latency connections.
I don't know why Most Wanted U's multiplayer is limited compared to other versions. Honestly I've never felt like the 6 vs 8 car difference ruined anything. It's still a blast.
blu
Wants the largest console games publisher to avoid Nintendo's platforms.
Ok, elaboration time.Krizz, net code and whatnot is run off the cpu. That is what we are talking about, no? Online multiplayer? I don't think the actual issue is rendering more cars.
Looking forward to what blu has to say regarding the TMUs. Just to clarify, except for some of the acronym identifications on Brazos, all thoughts/conclusions were my own. Tried to word my post carefully, but it doesn't hurt to restate that. I know everyone has their own thoughts.
Glad people enjoy the analysis. It's a fun passtime.
Disclaimer: I'm using only the high-res Brazos die-shot + detailed floorplan and the high-res Latte die-shot + 'unknowns' floorplan, as those provide sufficiently detailed pictures where I don't have to guess the block boundaries /disclaimer
First, a foreword: I do agree that determining the number of TMUs and ROPs can be used as a good guideline for determining the number of SPs. A quick look at R700 and R800 historical data reveals that for those two gens:
- SP-to-TMU ratio is in the set of { 10, 20 }-to-1
- SP-to-ROP ratio is in the set of { 20, 40, 45, 50, 70, 80 }-to-1
Where I disagree with you is your TMU count estimate, for nothing else but your block reading. First, I'm of the opinion the original 'bunch of unknowns' floorplan from Thraktor's OP is still perfectly valid - each block in that plan is actually a block. That includes the doubles. So what you call blocks S in my opinion is indeed two identical blocks S1 and S2 + some material from T1 and T2 (which while not identically laid out, are still apparent doubles).
- Status
Similar threads
- Poll
- Poll