Wii U CPU |Espresso| Die Photo - Courtesy of Chipworks

Feb 17, 2006
6,713
0
0
#1


High resolution image (at long last)!

What’s going on here?
This is the long awaited sequel to the WiiU "Latte" GPU Die Photo - GPU Feature Set and Power Analysis thread. If you have somehow missed it up until now, why not stop on over for some fervent technical discussion? If you’d rather not, here’s an abridged version of the events leading up: Desperately seeking some technical specifics in the face of Nintendo’s silence, a few GAF posters in the Wii U Technical Discussion Thread banded together in order to raise $200 for a die photo of the GPU from the wizards at Chipworks. Before the transaction was made, however, Chipworks contacted us with some pertinent info. It turns out that it would not be possible to discern the specifics we sought from the top silicon die shots offered on their site. What happened next was quite extraordinary. Chipworks went out of their way to take new polysilicon die shots of both the CPU and GPU (which normally sell for $2500 each) and have forwarded them to us, free of charge, for all to enjoy.

So What Is It?
Espresso is a tri core custom PowerPC design manufactured on IBM’s 45nm SOI process.

Vital Statistics:
  • Die Size: 4.74 x 5.85mm
  • Clock Rate: 1.243125 Ghz
  • 3 MB eDRAM L2 cache:
    Core 0: 512 kB
    Core 1: 2 MB
    Core 2: 512 kB
  • 32 kB L1 Instruction cache per core
  • 32 kB L1 Data Cache per core
  • Locked L1 Data cache DMA per core
  • Write Gatherer per core

Other Info:

  • Espresso offers hardware backwards compatibility with Gekko/Broadway, and is very likely based on the same PPC 750 core.
  • 2-way SIMD is done via "paired-singles."
  • Each core can retire 2 instruction per cycle (note: this is NOT the same as multithreading)
  • In Wii mode, only core 0 is active, and half of its L2 cache is locked

Comparison Dies:
Broadway annotated courtesy of marcan

Special Thanks:

Jim, Rob, Randy and everyone at Chipworks who has gone out of their way to help us here at Neogaf.

Also thanks to Thraktor, blu, Durante, wsippel, Schnozberry, Untalkative_Bunny, pestul, Azak, OryoN, JohnB, DeuceGamer, Deviljho, popstar, freezamite, phosphor112, and Dynamite Ringo Matsuri for helping to advance this effort.

Credit to marcan and vgleaks for additional CPU info.

(OP will be edited with new info as we go)
 

Wahooka

Neo Member
Sep 7, 2012
52
0
0
#4
As a person who doesn't fully understand what he is looking at I just want to say thank you to Chipworks for the image and for all of the other posters who will be dissecting it!
 
Sep 8, 2006
8,190
0
0
#10
This is what I was holding out for.

So I'm guessing the black blocks are the L2 Cache? I'm counting 12 making it 256KB per block for 3MB.

The center core seems to have some extras as well.
 
Oct 9, 2005
7,115
0
0
#11
This one looks pretty easy compared to the gpu die...

All the cores look to be the same and you can see the edram and larger cache of core 0 in the middle.

Not much hidden magic in this one.
 
Sep 8, 2006
8,190
0
0
#14
The center core only seems to have more SRAM for L2 tags as far as I can tell - because it has more L2. ;)
So those are the tags? Ok. Now there is also a small purple block between the tags and the L2 of the center core. And there seem to be a couple extra blocks on the center core above "bottom" core. One is green with white on it and the yellowish-green block right behind it. I don't see them on the other cores.
 

FLAguy954

Junior Member
Jan 27, 2012
5,155
0
0
USA
#21
So those are the tags? Ok. Now there is also a small purple block between the tags and the L2 of the center core. And there seem to be a couple extra blocks on the center core above "bottom" core. One is green with white on it and the yellowish-green block right behind it. I don't see them on the other cores.
And what about that purple block located in-between the top and middle cores?
 
May 25, 2006
14,534
0
0
Erfurt, Germany
#22
So those are the tags? Ok. Now there is also a small purple block between the tags and the L2 of the center core. And there seem to be a couple extra blocks on the center core above "bottom" core. One is green with white on it and the yellowish-green block right behind it. I don't see them on the other cores.
I don't think those blocks are actually part of the core.
 
Sep 8, 2006
8,190
0
0
#25
^ Are those the biggest ones you could find? *I couldn't even find that for the 750/Gekko*

And what about that purple block located in-between the top and middle cores?
You're probably talking about the same one I am.

I don't think those blocks are actually part of the core.
Probably. There does seem to be extra space to accommodate the latter two blocks I mentioned, though I would hazard a guess their position relates to center core. There are also some thinner blocks that aren't as visible down from the purple block.
 
Jul 3, 2004
9,938
0
0
31
Sweden
#26
In the left side middle top right area is what looks like a SARS-apparatus, probably functioning as a wideband digital amplifier resonance cache. It could indicate a 48nm inverted custom build. The right side lower left side blocks look like 9*9*9 randomly accesible memoriams, assuming they are neutral density of course.

Calculated guesswork gives me a terafloppy of 540 blue.
 

MDX

Member
Aug 11, 2010
2,330
0
0
#27


Akagi: I worked on the SDK and absolutely love CPUs! (laughs) So I myself have been happy to see how far game consoles have come. I hope people will be impressed by the types of software enabled by the CPU.

Iwata: I like the way you say you love CPUs! (laughs) It turned out to be a CPU without any strange habits—one that runs just the way you expect.

Akagi: (Laughs)
 
May 18, 2012
3,164
0
0
Mexico
#30
While I only follow these threads to a certain extent (not being able to post something meaningful means that I'm stuck at only watching), I really have to give mad props to ChipWorks for the amazing contribution that they did.

Seriously guys, you rock.

Now, back to lurking the thread.
 
Nov 20, 2012
1,794
0
0
#35
It's hard to make much out here. The eDRAM blocks are obvious, but I can't tell much else apart. I can't even tell the cores from the uncores right now (edit: ok yes I can, it's just not as stark a difference as usual. Three clusters of stuff to the right-ish, it does look like Core 1 has access to the middle eDRAM while the others don't?), and you could with any of Chipworks other CPU die shots, did they say why it's like that? Is it a flip chip package and that's why there may have been additional metal layers over the architecture making it hazy?

And you said each shot was 2500 dollars, I thought they were 200 dollars, or is this a different type of shot that we're looking at which is more advanced for whatever reason? Does it see stuff a regular one wouldn't, which would also explain the strange look?
 
Mar 29, 2012
10,988
0
0
Sweden
#36
I consider myself pretty "techy", but I don't have any knowledge on hardware at this level so I'll just thank everyone for their hard work and I will continue to follow these kind of threads.
 

Gahiggidy

My aunt & uncle run a Mom & Pop store, "The Gamecube Hut", and sold 80k WiiU within minutes of opening.
Jun 12, 2004
7,930
5
0
41
#37
Looks more "organic" than I would have expected from a computer chip.
 
Dec 22, 2008
2,817
3
665
#39
^ Are those the biggest ones you could find? *I couldn't even find that for the 750/Gekko*
Taking Broadway aside... Sadly yes; I left the sources for them linked to the images though; but for Flipper and PPC 750 the images are really that sized.


I actually suspect one of the sources could have a huge ass PPC 750 die image lying around; seeing they have this detail for it:



We spotted this medieval 300-micron long sword near the scribe line on a Motorola/IBM PowerPC 750 (...) the G3 chips were code-named "Arthur" as in Camelot, and the sword represents Excalibur.
Source: http://micro.magnet.fsu.edu/creatures/pages/g3sword.html


I've thought of emailing them for it, but haven't got to it. Probably won't, so if any one of you guys wants to play the diplomat go ahead. ;)



And since I did it, I'll leave this side by side thing here:




Feel free to use it if it's useful for the OP.
 
Dec 22, 2008
2,817
3
665
#42
And you said each shot was 2500 dollars, I thought they were 200 dollars, or is this a different type of shot that we're looking at which is more advanced for whatever reason? Does it see stuff a regular one wouldn't, which would also explain the strange look?
The 200 dollar shot would show us this:



Nothing more.
 
Dec 22, 2008
2,817
3
665
#45
Decapping and all that, done with the quality they're known for is really expensive. Understandably so.

As for the Espresso picture, I can identify the L2 cache blocks (middle core has 4 blocks so that's 2 MB worth; others have one so it gives it away)

L1 Instruction and data cache is also evident, although it seems doubled in size/banks compared to predecessor die shots; could we be dealing with 64 KB Instruction and 64 KB Data? Or did they opt for 2x16 KB? L1 tags and respective Translation Lookaside Buffers seem to be paired together and L2 Tags are also guessable from a number of hypothesis.

Upper middle part is probably the fuses I'm guessing. This going by marcan's Broadway anotation.


EDIT: Isn't there a bigger image for the chipworks die than the one sized for the OP?
 
Jul 21, 2007
11,843
0
0
Bavaria, Germany
#46
Thanks, Chipworks!

Also, thanks to Fourth Storm and Co for making this possible.

That said, my, what a difference to the GPU shot. Amazing.

We need a block overlay before we can really discuss anything. I shall now glare at the people in the thread and tap my feet impatiently. C'mon, get to it, guys! What am I paying you for, geez.
 
Dec 22, 2008
2,817
3
665
#49
Funny thing, assuming a straight Broadway core shrink to 45 nm (from 90 nm) would go from 18.9 mm² die area to half that, 9.45 mm².

Then 3x9.45mm2 = 28.35 mm² die size. Add more cache and SMP interface and it should go at least a few mm² higher.

But this chip, going by the OP seems to be 4.74 x 5.85 = 27.73 mm²; so less than expected despite having a huge area dedicated to the block things on the left and 4 times the total cache a 3 core Broadway would have. Sure buses like the 60x bus controller and the like wouldn't be tripled, but still; i'd expect it to not go lower than expected (I also realize I could have made this math before, I guess).

Going by that seems like more than a simple core shrink with rearranging, no? There's either shared components (mark of a properly planned out multicore processor) or something else.
That is just a bit misleading. Getting top silicon die shots is still a pretty involved process and the results look nothing like the one in your post.
I stand corrected.

But how would it look then; they would still decap it to an extent?