A Summary of the Huge Wii Thread

[Teasy]

I'll take the estimates from the first post:

106mm^2@180nm would be (106/4)mm^2@90nm which is:

26.5mm^2

and we have that the Hollywood has 72 mm^2, that is

(72mm^2)/(26.5mm^2) ~= 2.717

or in layman terms close to 3 times as big. Basically it should have at least about +150% more transistors.

I had thought that Flipper was 112mm^2 (was going to post a correction to your post) but looking around on the net it does seem like it is actually more like 106mm^2.

Anyway by this estimate Hollywood should be around 130 million transistors (assuming some inefficiency in the shrinking process). It seems pretty certain that the chip has the same amount of embedded memory as Flipper, which takes up about 25 million transistors. Which would mean that Hollywood has just over 100 million logic transistors vs 25 million for Flipper.

That does seem hard to believe considering what we've seen so far and what we're hearing about the chip (no vertex shader ect). But at the same time I can't see there being more embedded memory and the chip sizes are fact. So unless there's something nobody here can think of to explain away the differences in sizes it seems the chip must be quite a way more powerful then we've seen so far.

 

[DuckThor Evil]

So unless there's something nobody here can think of to explain away the differences in sizes it seems the chip must be quite a way more powerful then we've seen so far.

Maybe Dave Baumann will come here in disguise and leaks us the information? That would be nice wouldn't it.

 

[Capeta]

But at the same time I can't see there being more embedded memory and the chip sizes are fact. So unless there's something nobody here can think of to explain away the differences in sizes it seems the chip must be quite a way more powerful then we've seen so far.

Yeah seems pretty strange if it was a straight shrink and clock upgrade. If I had to guess I'd say there's more than 3MB of eDRAM and more logic deditated to more fillrate/geometry/lighting and maybe even some physics, no fancy VS though.

 

[Shifty Geezer]

Does anyone still believe it's just a die-shrink + clock-speed hike? I thought it was accepted that the architecture was probably something like a doubled-up Flipper; fundamentally the same architecture (eg. no conventional SM2/3 class programmable shaders) but running faster. That explains the use of the term 'overclocked' (improved speed with no feature improvements) and the difference in die size from expected.

 

[Capeta]

Does anyone still believe it's just a die-shrink + clock-speed hike? I thought it was accepted that the architecture was probably something like a doubled-up Flipper; fundamentally the same architecture (eg. no conventional SM2/3 class programmable shaders) but running faster. That explains the use of the term 'overclocked' (improved speed with no feature improvements) and the difference in die size from expected.

See the problem is the speculation could go in any direction since nobody knows what's in it. What if Hollywood is just a die shrink/higher clock and more eDRAM? That's certainly one valid reason why the die is bigger than it should be.

 

[fearsomepirate]

Paired single-precision arithmetic = 2/4 floating point operations per instruction on couples of single-precision floats held in the same registers. In my book this is called single-instruction multiple-data: SIMD.

Yes, it is. But there is a difference between "SIMD instructions" and a "SIMD unit." Gekko has an otherwise standard FPU capable of taking some limited SIMD instructions. It does not have a dedicated SIMD unit.

 

[pc999]

Does anyone still believe it's just a die-shrink + clock-speed hike? I thought it was accepted that the architecture was probably something like a doubled-up Flipper; fundamentally the same architecture (eg. no conventional SM2/3 class programmable shaders) but running faster. That explains the use of the term 'overclocked' (improved speed with no feature improvements) and the difference in die size from expected.

Even if you double the TnL, pixelpipelines and all that is doing gfx work that would give you a die 3x as big? If you consider that at least 1/3 of the flipper die is edram (and it should be, now, about 1/3 of that according to some rdram doc from the old thread) and part of the rest is I/O, sound DSP and such, would it end up it more than a 50% bigger die than Flipper? That is still almost 1/2 of what we see here.

 

[Ooh-videogames]

When did devs get final hardware?

I still wonder if the maxconsole specs were based final hardware. Devs just won't reveal anything lately. Maybe it is just double everything.

 

[Urian]

Hollywood continues being Flipper, perhaps the extra space are coprocessors that helps to do some complex graphical effects that supposes a critical hit in the performance of the GPU.

 

[Corwin_B]

See the problem is the speculation could go in any direction since nobody knows what's in it. What if Hollywood is just a die shrink/higher clock and more eDRAM? That's certainly one valid reason why the die is bigger than it should be.

More EDRAM doesn't seem likely, though. First, 3MB of EDRAM made around 1/3 of the Flipper die, so tripling die size (Flipper -> Hollywood) would mean an extra 18MB of EDRAM. That would be totally overkill for a SD chip.

 

[ERP]

That does seem hard to believe considering what we've seen so far and what we're hearing about the chip (no vertex shader ect). But at the same time I can't see there being more embedded memory and the chip sizes are fact. So unless there's something nobody here can think of to explain away the differences in sizes it seems the chip must be quite a way more powerful then we've seen so far.

I'm not confirming anything honestly this is far too much wishful thinking and what de-railed the original thread IMO.

Things don't stay secret. People just selectively choose what to believe.

Both parts have to have additional memory controller logic to deal with the GDR memory, it could be that Nintendo wanted to add redundancy to improve yields. (I have no idea), or any number of none functional reasons.

In general Devs will be doing what they can with what they have, not holding back for some second gen.

 

[MDX]

Ok, I never want to do that again. I need to forget that thread or I'll have nightmares about it... Another thing I want to point out beforehand: this is just the bare summary and I tried to stick to the "facts" as close as possible

I'll start off with the hard facts (100% confirmed), followed by the nearly 100% confirmed (if possible I'll try to cite multiple sources). So let's start:

A look inside Wii

Wii dismantled:
http://pc.watch.impress.co.jp/docs/2006/1201/nintendo.htm
http://pc.watch.impress.co.jp/docs/2006/1129/nintendo.htm

Nikkei article (thanks to Mmmkay)


compared The "Gekko" (CPU) in the first GameCube occupied 43mm2.@ 180nm
Flipper 110 mm2@180nm


Schematics
Scheme

source Goto article on future revisions; see also http://www.beyond3d.com/articles/futureconsole/


Memory:
64 MB GDDR3 + 24 MB 1T-SRAM + embedded RAM (most likely 3 MB like Flipper)
Specs of the GDDR3 (found by zeckensack)




Broadway:
Official:
SOI by IBM
90nm

Goto:
http://pc.watch.impress.co.jp/docs/2006/0920/kaigai301.htm
* "Broadway based on Gekko" & customized 750CXe; theafu claims to have insider knowledge and it's a 750CL LINK; my money is on Goto, though ;-)

Maxconsole Net
Maxconsole Net's specs
Note: the maxconsole net specs have been backed up by at least 2 other sources (
Different source backing up the Max Console specs).



Hollywood
Official:
Hollywood designed by ATI
90nm
embedded DRAM
LSI

Goto:
http://pc.watch.impress.co.jp/docs/2006/0920/kaigai301.htm
* GPU lacks shaders -> TEV texture combiners

Maxconsole Net


Engadget (Comments section)
Different source backing up MaxConsole


Matt from IGN also pointed to no shaders (Disclaimer: It's Matt!)


so far we've also seen dithering problems in Wii titles -> GCN legacy 24bit incl. alpha???. See post here



about the TEV
Old GCN article on the features of the TEV
old thread on TEV with comments from ERP
Ingenu post with old Flipper patents incl TEV

Wii/console/motherboard:

top

bottom

Chip listing:
U1: "(E)" "HOLLYWOOD" "(C)(M)'06Nintendo" "(ATI logo)" "(BroadOn logo)" "(NEC logo) JAPAN" "010046F5-211-PN2-A" "0633KK01X" (video/IO chip, BGA/unknown 31x31mm)
U2: "(IBM logo)" "BROADWAY" "39X6735" "IBM9316" "BWYC-72914" "06340706V 14" "CANADA" (CPU, BGA/unknown 21x21mm)
U3: Samsung K4J52324QC-BC14 "SAMSUNG 631" "K4J52324QC-BC14" "EFEA75DX" (Memory!, BGA/unknown 14x11mm)
U4: Sharp PQ070XH02Z "070XH02" "SHARP" "12H U6" (LDO voltage regulator, set at 1.8V - TO-263 package, 5 pins 10x8mm)
U5: "H8DU" (*tiny* 5-pin tsop-6 3x1.5mm)
U6: "AVE-RVL" "BU9055EKV" "629 145" (tqfp-64 10x10mm)
U7: Mitsumi "616" "AMPAM" (tsop-8 5x4mm)
U8: MX "E062674-MG" "MX23L4005-24C1" "RTC-RVL A" "2S08802" (RTC, TSOP28 18x8mm)
U9: "JK0U" (Power-on-reset generator *tiny* 5-pin tsop-6 3x1.5mm)
U10: "JK0U" (Power-on-reset generator *tiny* 5-pin tsop-6 3x1.5mm)
U11:
U12:
U13:
U14: Samsung K9F4G08U0A "SAMSUNG 631" "K9F4G08U0A" "PCG1" "FHFL82GAU" (flash memory, tsop-48 18.5x12mm)
U15: Sanyo LV5043 "LV5043" "6SP0" (DC-DC Converter controller, ssop-30 10x5mm)
U16: Sanyo LV5043 "LV5043" "6SP0" (DC-DC Converter controller, ssop-30 10x5mm)
U17: Sanyo LV5043 "LV5043" "6SP0" (DC-DC Converter controller, ssop-30 10x5mm)
U18: "2901" "6M56" (tsop-14 5x4mm)
U19: Sharp PQ033DNA1ZPH "033DNA1" "SHARP" "H U7" (LDO regulator, 3.3V 1A, 4 pins 6x5mm)
U20:
U21:
U22: "H4" (*tiny* 5-pin tsop-6 3x1.5mm)
[edit]
Ports
P1 Gamecube memory cards port
P2 Rear fan power connector
P3 Disc slot LED port
P4 Disc drive port (see Wii console Disc Drive page)
P5 AV and "sensor bar" port
P6 Gamecube controller ports (all 4 of 'em)
P7 Battery port
P8 Disc drive port (see Wii console Disc Drive page)
P9 USB ports
P10 SD Card port. Marked "A0635D6"
P11 Power connector
P12 Bluetooth Daughtercard
P13 802.11 Wifi daughtercard
[edit]
Crystals
X1 "6373" "T 6276"
X2 "32.768" "KDS0631" (32768Hz, RTC oscilator)

http://wiire.org/Wii/console/motherboard

 

[Teasy]

ERP

Expecting DX9 pixel/vertex shaders would be wishful thinking at this point. Expecting extra pixel pipes, T&L unit ect seems logical considering the size of the chip.. Do you think any of the suggestions you've made to explain the chip size are actually more likely? (serious question)..

Also yeah I'm sure devs will be doing what they can and nobody thinks any of them are purposly holding back to try to suprise people with super second gen Wii games. But the fact is some developers are more skilled then others and/or simply place more focus/effort into graphics (especially on a console where getting to grips with a new interface is the initial priority). The fact that the majority of Wii games so far have been no better (in fact mostly worse) looking then GC games enphasises that pretty well, and is echoed in Julian Eggebrecht's recent comments.

 

[crystall]

Yes, it is. But there is a difference between "SIMD instructions" and a "SIMD unit." Gekko has an otherwise standard FPU capable of taking some limited SIMD instructions. It does not have a dedicated SIMD unit.

Actually every single precision floating point instruction is supported in SIMD mode, on top of that additional instructions are available in paired mode, not just 'some limited instructions' on top of that L/S instructions are available to fill/spill the FP registers with couples of SPFP values either as is or converting them to/from 8/16-bit fixed point. Now if by 'some limited instructions' you meant that it doesn't have integer SIMD capabilities then this is absolutely correct but that doesn't mean that Gekko lacked SIMD capabilities of any sort.

Aside from that by your definition none of the x86 processors available today has a 'dedicated SIMD unit' as they all share one of the SSEx FP MUL and ADD pipes with the x87 FP unit.

 

[Capeta]

More EDRAM doesn't seem likely, though. First, 3MB of EDRAM made around 1/3 of the Flipper die, so tripling die size (Flipper -> Hollywood) would mean an extra 18MB of EDRAM. That would be totally overkill for a SD chip.

According to NEC, at 90nm 32MB of eDRAM takes up 225mm^2 that means for every 1MB of eDRAM you'd need roughly 7mm^2. There's no way you could fit 18MB of eDRAM into 70mm^2 + Flipper logic. Actually 3MB of eDRAM takes up 21mm^2 at 90nm which is exactly 30% the area of Hollywood which means it IS indeed an overclocked Flipper.

So there you have it, I guess we can stop the nonsense now. Hollywood IS a simple die shrink/clock increase.

 

[fearsomepirate]

So there you have it, I guess we can stop the nonsense now. Hollywood IS a simple die shrink/clock increase.

That's only if the rest of Flipper shrunk in the same proportion as the eDRAM, which I suppose is possible, since they'd need to keep all the connections where they are. But it certainly seems odd that moving from a 180nm to a 90nm process only reduced die size by 32%.

 

[Ooh-videogames]

Is that ed-ram the 1t-sram used on flipper, or the 1t-sram-r which is supposed to be manufactured on the 90nm process?

 

[ATI-liens]

Is that ed-ram the 1t-sram used on flipper, or the 1t-sram-r which is supposed to be manufactured on the 90nm process?

The 24MB 1T-Sram is 6ns, so i doubt the ed-Ram will be the same as the NGC.

 

[The_legend_of_drtre]

According to NEC, at 90nm 32MB of eDRAM takes up 225mm^2 that means for every 1MB of eDRAM you'd need roughly 7mm^2. There's no way you could fit 18MB of eDRAM into 70mm^2 + Flipper logic. Actually 3MB of eDRAM takes up 21mm^2 at 90nm which is exactly 30% the area of Hollywood which means it IS indeed an overclocked Flipper.

So there you have it, I guess we can stop the nonsense now. Hollywood IS a simple die shrink/clock increase.

The GC's flipper already had 3MB of dram. Hollywood is now on 90nm and is still much bigger than flipper. It was not a die shrink and clock increase. Add to the fact that now the 24MB of 1t-sarm is now located inside the chips die instead of outside of it like on flipper. They are not the same chip.

 

[Capeta]

Is that ed-ram the 1t-sram used on flipper, or the 1t-sram-r which is supposed to be manufactured on the 90nm process?

It's their newest low power high density 90nm eDRAM process so it's not 1T-SRAM-R.

http://www.necel.com/process/en/edramoptions.html

Look at the table and you can see the cell size comparison of the NED3 process vs UX6D process.

http://www.necel.com/process/en/edramprocess.html

The GC's flipper already had 3MB of dram. Hollywood is now on 90nm and is still much bigger than flipper. It was not a die shrink and clock increase. Add to the fact that now the 24MB of 1t-sarm is now located inside the chips die instead of outside of it like on flipper. They are not the same chip.

First of all what does the 24MB of 1T-SRAM have to do with anything? Second, according to NEC's numbers Hollywood is a straight die shrink/clock increase since 30% of the die is used for the 3MB of eDRAM. This is the low power higher density eDRAM process we are talking about too, not their low density high performance one. Believe what you want to believe, but when Wii games don't show significant improvement over GC, don't pretend you didn't hear it here first.