Aizu_Itsuko
Member
I want to thank Chipworks as well for helping us shed some light on Wii U's hardware.
Really appreciated.
Really appreciated.
-
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.
Wii U CPU |Espresso| Die Photo - Courtesy of Chipworks
- Thread starter Fourth Storm
- Start date
grimshawish
Banned
Wow. This is amazing of them. Thank you chipworks.
blu
Wants the largest console games publisher to avoid Nintendo's platforms.
There's nothing gimped about it - it's just small-scale. The 750 does not have as many units as the 970 so there's no point in it resolving 5 ops for dispatch, and the entire associated workload. But 750's design performs notably better than an in-order design, at a minimal transistor premium.
if the cpu is truly an FX and if the middle core is more bull boosted core, that means is dedicated for the games ? Definitely a better geek than xbox360.
360 uses 1 core 2 threads for the games, 1 core for the sound and 1 core for the OS. I am satisfied with the sounding of that... Better gpu and better cpu, it only needs different coding and more attention from the developers.. Wii U cpu (750 family) may be single cores (no threads), but they perfom with more instructions per clock, than any other... I think 360 is about 5 instructions per clock but a broadway is about 7+????
edit: How many instructions per clock a 750FX gives? Broadway gives 7+ (7.2 if I am right)
360 uses 1 core 2 threads for the games, 1 core for the sound and 1 core for the OS. I am satisfied with the sounding of that... Better gpu and better cpu, it only needs different coding and more attention from the developers.. Wii U cpu (750 family) may be single cores (no threads), but they perfom with more instructions per clock, than any other... I think 360 is about 5 instructions per clock but a broadway is about 7+????
edit: How many instructions per clock a 750FX gives? Broadway gives 7+ (7.2 if I am right)
frankie_baby
Member
dont know how much the 360 os uses but sound yes can use a whole core though usually just one thread, though yes for general code (not floating point) the wii u should have noticably more cpu power available to it than 360
Is that accurate? I know it uses 1 thread for sound, and "can" use up to 1 core, but remember hearing only a handful of games needed 1 entire core. And does it use 1 core for OS while gaming?
I ve read in an article, in the past, even skyrim uses 1 core, I know that because it took almost 6 months till they patch it properly and perform well in the PC. Also if you measure ports from 360 games in the pc, they always use 1 core.
XBOX 360 can use the OS during gaming, by pressing the middle green button, also its not only the game sound, but the cross chat, etc.. It has many features, downloading in the background, etc...
This is from Anandtech : http://www.anandtech.com/show/1719/6
With Microsoft themselves telling us not to expect more than one or two threads of execution to be dedicated to game code, will the remaining two cores of the Xenon go unused for the first year or two of the Xbox 360’s existence? While the remaining cores won’t directly be used for game performance acceleration, they won’t remain idle - enter the Xbox 360’s helper threads.
edit: I hope I helped. It also says, due to low dvd capacity, xbox uses 1 thread for decompressing data.
XBOX 360 can use the OS during gaming, by pressing the middle green button, also its not only the game sound, but the cross chat, etc.. It has many features, downloading in the background, etc...
This is from Anandtech : http://www.anandtech.com/show/1719/6
With Microsoft themselves telling us not to expect more than one or two threads of execution to be dedicated to game code, will the remaining two cores of the Xenon go unused for the first year or two of the Xbox 360’s existence? While the remaining cores won’t directly be used for game performance acceleration, they won’t remain idle - enter the Xbox 360’s helper threads.
edit: I hope I helped. It also says, due to low dvd capacity, xbox uses 1 thread for decompressing data.
That is not right. Developers have access to every core and thread on the 360, one *thread* has to be used for sound and in games with a lot of sounds it may take that whole core, but can often just use one thread of one core. The OS does not reserve a whole core while gaming. At the very most you could say 1 core may be tied up with sound, but devs have full access to the other two with four threads total.
By your logic, if one core was used for sound and one for the OS, adding Kinect into the mix would leave zero cores for games
(kinect by the way uses a single digit percentage of the whole processor, I imagine it's similar with the OS, it's on just enough for a few functions)
What does "even skyrim" mean, that's their fault, not the 360s. Skyrim had poor CPU distribution regardless of platform.
blu
Wants the largest console games publisher to avoid Nintendo's platforms.
Depends what your definition for instructions-per-clock is. According to the common convention for IPC, both Xenon and Broadway are 2 IPC (per core).
lostinblue
Banned
I'm not doubting you, but you should doubt that story.
2 GHz between 2003 and 2006 for a PPC 750 part wouldn't be possible without messing with the pipeline in a big way; pretty much to the point of turning it into a G4. It's like saying they had gold but didn't release it and that simply doesn't cut it (it wasn't Apple alone that used PPC 750's, routers and embedded devices could benefit from it)
We have to be realistic; the last G4 released wasn't budging past 2.0 GHz in 2005, sure it was Freescale, but IBM had heat problems of their own in a pretty big way too, and their solution couldn't be going for a 6 staged architecture. PPC 476FP only hit 2 GHz in 2009 or so and it's 9-staged (PPC750 is 6 stages), e500, which is PPC 750 spiritual sucessor only hits 1.5 GHz today.
1.25 GHz Power hungry G3 I can bite, as it's only 150 MHz over 2004's PPC GX but 2 GHz in that timeframe is pure nonsense, that totally gives it away, it's an exagerated myth.
Also, G5 performance was not poor by any means, it was just not ready for laptops and the companion chipsets for it were very inefficient (energy and performance-wise, memory speed was hampered by it), but it still managed to perform really well.
I was also hard to scale up. Saying they had a G3 @ 2 GHz is like stating they had a PPC 970 @ 3.2 GHz without extreme cooling before 2006 or so. Clearly impossible.
And I don't get this part, is he saying xenon and cell are ppc750 based or that g4's are ppc 970 now; or... is he saying that cell and xenon architectures were originally meant for general purpose? Either way, false for all counts.
Also entertaining, he thinks intel *supposed* alliance with apple created the iCore (i3/i5/i7)?
You are making some broad generalizations that really have no foundations.
First you cannot compare PC processor usage of any indication of what the Xbox cores are doing.
Second, you are confusing a cores "thread" and a task thread. You can have many times more task threads than physical threads. In fact hit ctrl+shift+esc and look at all the threads your windows box has going at any given time. It's no difference with the 360 or any modern operating system (embedded or not). The CPUs threads are just what the CPU can physically deal with at a given time, there is a scheduler to handle more virtual threads than physical ones. If you have something critical, you *may* want to dedicate a physical thread/core to its duties, but its not a requirement. MS I believe does not limit the developer to give up a core to the os, Sony does keep one spe for os duties though.
*edit* I just scanned the attached article,
I dont know how relevant it is to 360 games of today. It was written at the launch of the 360 before even pcs were heavily multithreaded. An interesting read though.
So far, besides pointing out the probable cache locations, all I've seen on this thread is people trying to defend the 360 CPU, insist the Wii U's is just 3 Broadway CPUs taped together, and dismiss any positive performance allotted to the PPC750 architecture.
When are we going to actually start doing the Espresso analysis?
EDIT: Also, I've been wondering about the performance benefit of the CPU and GPU being on the same die. How does that work? What exactly is the benefit as compared to having them separate and independent?
When are we going to actually start doing the Espresso analysis?
EDIT: Also, I've been wondering about the performance benefit of the CPU and GPU being on the same die. How does that work? What exactly is the benefit as compared to having them separate and independent?
frankie_baby
Member
I'd be more than happy if it was 3 broadways duct tapped together as ive been in the camp that thats a good thing anyway, if there is more then id be very happy chappy
Schnozberry
Member
Seems there are a lot less unknowns here. It isn't some hyper customized PPC750, aside from the cache customizations.
Seems like a very good fit for what Nintendo was trying to do. Along with the well designed memory subsystem and modern GPU, the system should, as Criterion said, punch above it's weight. Doesn't mean we should expect miracles, but we should get some very impressive looking games.
Now, it's just up to Nintendo to reverse the staggering loss in sales momentum. We need games and a new marketing plan ASAP, otherwise it could be a while before Nintendo returns to profitability.
This is completely unrelated, but how weird is it to be a fan of an industry where the two platforms that lost a combined 8 billion dollars on hardware over the last 8 years are considered to have the most positive momentum. Gaming as an industry could he in for major tremors over the next few years.
frankie_baby
Member
As someone asked on the previous page if anyone could resize the Broadway and/or espresso pics to the size they'd be on the same process that would really help, can't do it myself as my PC is in pieces and I'm browsing on my Wii u
It's a good question, though to clarify, you mean CPU/GPU on the same MCM (multi chip module package), they're not on the same Die (which would be silicon). Apples A[x] series on the other hand have the CPU and GPU on the same die.
But my similar question would be what would be the benefits of putting them on the same MCM over having them just on the same motherboard? At what point does an MCM just become a motherboard anyways? Would it be theoretically possible to put the main memory on the MCM as well, or would that just make the MCM a motherboard. Is the MCM similar to a mainboard but with the potential for higher speed/wider bus' between components? I would guess that the MCM would also have a different fabrication/assembly process.
How are sales relevant to the CPU capabilities or the games "looks"? I do not follow.
On what grounds do you come to the conclusion that it is not extensively customized?
I've heard a few comments suggest that the performance we are seeing from the Expresso would be unlikely to be achieved on simple Broadway architecture. We cannot simply assume that because they are of the same family that they are also of relative performance, ie. an overclocked Braodwayx3.
None of what I am hearing is making sense to me. The points keep contradicting.
I wish I was more versed in component design. The 3 sections to the right have been identified as the cache. What are the 2 long black areas to the left and the 9 rectangles that I see towards the top? Also, are those 2 squares lasso components under it or something irrelevant? I should probably leaves this up to the pros.
I place Broadway under it for the guy who wants them the same size. I don't know how to place them next to one another.
.
EDIT: I decided to use Marcan's annotaed Broadway from the OP.
When people talk about the possibility of Espresso being three Wii CPUs "taped together" it really has to be remembered that at minimum we know theres much more L2 cache and that its clocked 71% faster. Safe to say Espresso's minimum performance should be around about 6 Wii CPUs.
Hey, do you know which is the Video Decoder Core?
I work on the software for PowerVR VXD and it'd be cool to see it actually on-die.
prag16
Banned
With the "punch above it's weight" comment, I wonder if it means in terms of power consumption, or in terms of raw numbers.
Gamecube punched above it's weight in both categories, so hopefully with the Wii U it's the same deal.
In theory 90nm to 45nm would make it a quarter of its original size. But in reality die's never shrink perfectly, especially the logic parts (memory shrinks much more linearly and not too far off the theory I think). Gekko was 43mm2 so a perfect shrink from 180nm to 90nm should have meant 10.75mm2 for Broadway, when in reality its 18.9mm2. Though as you say Broadway wasnt just Gekko on a 90nm process, it did have some extra transistors in there.
While Espresso does have 12x as much L2 cache as Broadway (4x per core average as you say) its using eDRAM which uses far less space than the SRAM used for Broadways L2 cache. So you're only actually looking at about 4x as many transistors for the 3MB in Espresso as was used for the 256KB cache in Broadway. Since memory shrinks very close to 4x going from 90nm to 45nm the L2 cache of both CPUs should take up a very similar die area.
I don't know for sure, but here's where it is on the Tegra 2, that obviously doesn't use PowerVR GPUs but the video decoder is a separate block
I don't know for sure, but here's where it is on the Tegra 2, that obviously doesn't use PowerVR GPUs but the video decoder is a separate block
I see, thanks.
Funny to see that the Video Decode is a lot bigger than the Encode, given that Encoding is a lot more computationally intensive. Then again, Encoders rarely require as much format support as Decoders.
Maybe I can ask around the office if anyone can identify for sure the Decoder on that A6 die.
lostinblue
Banned
I see.
I thought it was the opposite actually, that cache was the most non-linear part while shrinking; seeing just because it's embedded doesn't mean it's done on the same process/node.
Seems like that's not the wildcard; and if it's non-linear then little can be gained from comparing die areas then.
Yes, the change in sizes really confused me at first, I thought the L2 tags were the SRAM and the left side blocks were something else. Definitely overcomplicated things.
Interesting insight on the transistor math, helps put things into perspective.
Fourth Storm
Member
I believe marcan when he says that it's currently completely off limits to developers. It's definitely in there primarily for BC. Perhaps Nintendo are also using this as a cache for their own system features? It's always possible they open it up to devs in the future.
The CPU should be well fed as it is w/ the WiiU memory hierarchy. A bunch of L2 cache on chip, quick access to to the 32MB eDRAM, and the DDR3 as a last stop.
I hope it is opened up later, but I suppose it's possible it's only physically wired to the backwards compatibility bits. Extra bandwidth/cache levels for the GPU or CPU to work with can't be a bad thing, and the two pools that are unavailable are presumably faster than the 32MB (one for being SRAM, one for being more tightly packed eDRAM which means lower latency).
Fourth Storm
Member
Well, marcan claims some stuff runs through the shaders in BC mode. This doesn't surprise me as TEV was basically a very early pixel shader, and from the start, blu said that BC should be achievable using modern hardware as long as they included that texture cache. I can imagine they've modified the TMUs as well to support Wii's nuances. So my belief is that whatever they're wired to for BC should be the same stuff that drives Wii U games.
Cow Mengde
Banned
I'm actually more interested in knowing what the Wii can do. Maybe we can finally get some real specs like polygon count/cpu speed on Broadway and Hollywood.
Fourth Storm
Member
I'm pretty sure real clock speeds are readily available. As for the polygon numbers, multiplying the Cube specs by 1.5 should do. Same geometry engine - just higher clock. Anything else they changed was very minor. Somebody mentioned a few new instructions for Broadway perhaps? I gotta look that up again...
I am not comparing any pc with 360! Its old news that 360 uses 1 core 2 threads for its games, why you dont search. Maybe the new releases, like crysis 3, use 3 threads, but I find it extreme. Here is also a small example with xbox720. We all read the rumors that it comes with 8 cores, but the new kenect will use most of the cores for it!!! So 360 have lots of things to maintain at the same time, as it does have kenect and it does drain recourses. Normaly is 1 dedicated core and it gets some help from other threads, about 10-15%, wrong??? Thats an example.
""" lthough the sensor unit was originally planned to contain a microprocessor that would perform operations such as the system's skeletal mapping, it was revealed in January 2010 that the sensor would no longer feature a dedicated processor. Instead, processing would be handled by one of the processor cores of the Xbox 360's Xenon CPU. According to Alex Kipman, the Kinect system consumes about 10-15% of the Xbox 360's computing resources. However, in November, Alex Kipman made a statement that "the new motion control tech now only uses a single-digit percentage of the Xbox 360's processing power, down from the previously stated 10 to 15 percent." A number of observers commented that the computational load required for Kinect makes the addition of Kinect functionality to pre-existing games through software updates even less likely, with concepts specific to Kinect more likely to be the focus for developers using the platform"""" So 1-9% is dedicated to the kenect on 360, also if you search how the sound works on 360, its all about decompression, because it does not have dedicated sound chips.
Lets focus on wii U cpu now, I will wait to see analysis.
I do want to get back to the Wii U, but you can't just drop that here and expect no one to address it. Saying Kinect uses 10-15% (edit: down to single digit percentage by your own source) of the CPU does not prove your assertions that only one core is used for game code. Even if that is true for some poorly multithreaded games, it's not a hardware limitation, the OS does not reserve one physical core as your earlier post stated. One thread up to one core can be used for sound, but developers can divvy things up however they choose, games with less sounds can use one thread alone, not one whole core. The OS just stays on for basic functions during gameplay and does not need an entire core for it.
http://msdn.microsoft.com/en-us/library/windows/desktop/ee416321(v=vs.85).aspx
http://geekswithblogs.net/jwhitehorn/archive/2007/01/27/104731.aspx
It may be true that some earlier games only used one core during the transition to multicore processors, but that is not a limitation in the system, it's a programming thing.
And the Anandtech article you posted said itself that it would probably just be for the first year or two, which means it's a programming problem, not the system reserving one core. And the core being used for real time decompression is optional, they can decompress one level at a time and free the core.
Really, because last I heard 3 of 8 was not "most", and that three is for the OS as well as Kinect. That doesn't imply anything for the 360, which uses a much less fat OS than the Duragno is rumored to.
So you saying that games on xbox use about 3 cores for the games, because the sound is real time and its not only the sound that runs with the game! I said that for the most modern games, but its not a certain thing. And you know that microsoft did that from the beginning, because the cpu is "in order execution" much slower. So a wii U (lets say FX model core, I will wait for gaf analysis) can easily perform like 2 threds+ from 360, with the appropriate coding, with 1 single core.
If i am wrong, please explain, why not.
edit: I am asking that, because I am reading and I am learning, I am not a spec guy.
If i am wrong, please explain, why not.
edit: I am asking that, because I am reading and I am learning, I am not a spec guy.
Sorry, I'm getting really confused by what you're trying to say. The speed of the cores is irrelevant, nor did I say 100% of all three cores is used for game logic, of course at least one thread has to be used for audio and some percentage has to be used for background processes. I'm just saying that the OS does not lock up one core like you seemed to imply earlier, if games in the first two years only used one thread that's because of the old programming models, it's not a hardware limitation.
The Anandtech article you linked as your proof says it itself, games were expected to only use one thread for logic in the first year or two, which directly implies they are not locked to that in any way, as multithreaded programming became more common they were able to better use more cores and threads.
I never debated whether the Wii U can run all the game code the Xenon can run. My only point from the start was correcting what appeared to be you saying only one core is left for games on the 360. Nor is in-order always a huge detriment by the way, the AT article explains all this, although the Xenon is inefficient for other architectural reasons.
lostinblue
Banned
Actually suspected as much, one of the reasons why I left the source linked in:
It's supposed to be an "official" die shot; except they were using a regular PPC 750 sans L2. As for paired singles I wouldn't know how that looks if it sat on my lap. But I'm guessing it adds some logic like seen on Broadway.
He's right and that sorts it out.It was bugging me tbh. I tried annotating the gekko die, comparing it to the broadway he did, and I obviously couldn't get past the fact there were no L2 blocks in there. Wondered at first if they were smaller or something; that's why I looked for a PPC750 die, only to confirm that. Then I just posted it anyway because visual aspect of dies is really not my thing (I take everything I think regarding it as a big grain of salt, so I disregarded it) and because I already had the work looking for those, thought to me: why not.
Means we have no proper Gekko die shot though, that would be interesting to see.
I think he just means he's sick of having to tweet things to gaf instead of being able to actually post here.
lostinblue
Banned
Yeah, I replied there (first time posting on twitter) and the character limit sure isn't funny.
Where exactly are the processors 3 Cores?
Also, are you certain that the entirety of the squared area to the right in the photo is pure cache?
https://lh5.googleusercontent.com/-...JI/AAAAAAAAAdw/FyuvSjP1Phk/s640/Espresso2.png
Also, are you certain that the entirety of the squared area to the right in the photo is pure cache?
https://lh5.googleusercontent.com/-...JI/AAAAAAAAAdw/FyuvSjP1Phk/s640/Espresso2.png
The area in yellow squares ARE the cores. [edited] The more solid looking square blocks to the further right and to the left of the middle core are l1 cache and the leftmost is the IBM eDRAM.
blu
Wants the largest console games publisher to avoid Nintendo's platforms.
If you're talking of the SRAM arrays on the right-most of the die - those are all L1 - 32KB data + 32KB instruction caches, per core.
Is that right? I thought the solidly colored smaller blocks to the right of each core and the left of the middle core are eDRAM, I don't know what the far right is but it doesn't look like any eDRAM I've seen before.
Then again the whole shot is kind of weird looking, almost like it was looking through a layer, seems more hazy than every other chipworks shot.
Edit: It's kind of shaped like IBMs eDRAM promotional images, perhaps you are right. And considering the size of the smaller blocks, I was probably wrong, it's probably just sleep deprivation talking
Oh I see. I originally was going to ask if everyone was sure the areas on the right were cache and not the Cores because I thought someone said they cache were earlier in the thread.
That clears things up.
So that leaves the 9 rectangles to the north, the two squares under it and that small T shaped thing above core 2. I don't suppose the small object towards the upper left is anything important.
Though, even given that we don't have those identified, I expected a CPU to have a lot..."MORE" on it. I guess I'm to used too looking at GPU shots. Now that I can see the cores though, I must say that they all look very different from Broadway.
Looking at the Broadway picture, I'm guessing the 9 rectangles to the top are fuses. The fuses in the lower left corner of Marcan's annotated Broadway picture have a similar layout. Though, I'm probably wrong. The fuses on Broadway area lot more uniform.
That clears things up.
So that leaves the 9 rectangles to the north, the two squares under it and that small T shaped thing above core 2. I don't suppose the small object towards the upper left is anything important.
Though, even given that we don't have those identified, I expected a CPU to have a lot..."MORE" on it. I guess I'm to used too looking at GPU shots. Now that I can see the cores though, I must say that they all look very different from Broadway.
Looking at the Broadway picture, I'm guessing the 9 rectangles to the top are fuses. The fuses in the lower left corner of Marcan's annotated Broadway picture have a similar layout. Though, I'm probably wrong. The fuses on Broadway area lot more uniform.