GoForce 5500: A 120mm² handheld GPU?

Arun

Unknown.
Moderator
Legend
If this is accurate... WTF. This is bigger than the GeForce 7300GS!
http://www.nvidia.com/page/pg_20060207466624.html

The 3D performance looks OK, slightly above that of MBX most likely, the featureset isn't anywhere near that of SGX though. On the other hand, it has tons of insane stuff like H.264 decode/encode @ 30FPS up to D1 resolutions, and 10MPix camera support (omfg, I want that on my mobile!). Still... Between 100mm² and 120mm²... God!

Hopefully NVIDIA just put those transistors into totally separate features, and they got some kickass power management in it so disactivated parts don't take any power. It's interesting anyway, considering it's a discrete chip which can do a lot of stuff that would otherwise be in a SoC - so I imagine using a "basic" SoC that mostly any handheld manufacturer has available, and adding this NVIDIA chip, could give you a ready-to-sell mobile in just a couple of months. As such, that's interesting wrt avaibility (i.e.: this is likely to be available much before SGX or Mali2, or Bitboys' latest stuff).


Uttar
 
Mortimer said:
Package size, not core size.
At least that's how I read the specs.
That'd make sense - anyone can issue a guess as to die size then, though? :)

Uttar
 
Why does that one look to me like a revampled AR1x-whatever? Geometry processors and pixel shaders are checked on both 4800 as 5500.

I have severe doubts that that's NVIDIA's idea of their true second generation PDA/mobile offerings. Fancy register combiners aren't pixel shaders in my book *cough*
 
Uttar said:
That'd make sense - anyone can issue a guess as to die size then, though? :)

Uttar
No :D
But there's a good 32M transistors in there just for the embedded SRAM. Logic dictates that the amount of logic would be at least in the same ballpark on a somewhat balanced design. Pun intended.

How much of an overstatement is "programmable pixel shading engine" this time? They claimed the same for the previous generation and ... meh. Haha.
 
Ailuros said:
Why does that one look to me like a revampled AR1x-whatever? Geometry processors and pixel shaders are checked on both 4800 as 5500.

I have severe doubts that that's NVIDIA's idea of their true second generation PDA/mobile offerings. Fancy register combiners aren't pixel shaders in my book *cough*
AFAIK AR10 is RivaTNT level pixel processing.
"Register combiners" in any relation to NVIDIA implies at least a Geforce 1 level of flexibility.

Unless someone can point me to the extension specs (their OpenGL ES implementation should expose whatever the chip can do, so they claim), I expect the worst case.
 
Ailuros said:
That's the estimated size for 8MB?
No, it's for the 640kB of embedded SRAM that's mentioned in the tech specs Uttar has linked to.
Old fashioned SRAM needs six transistors per bit. So that's 640k*8*6 transistors and I assumed they meant 1024, not 1000, when they wrote "K".

edit: OT: I'm totally allergic to IRC :-?
 
Uttar said:
That'd make sense - anyone can issue a guess as to die size then, though? :)

Uttar
Unless the chip comes in a Chip-Scale-Package (not likely due to stacking in this case), chip size and package size are more or less unrelated.

In 90nm, the 640 Kbyte SRAM alone should take somewhere between 15 and 30 mm2 (the 15 mm2 low bound is a rough estimate of what the Intel 90nm process is capable of; most non-Intel foundries have traditionally had much larger SRAM cell sizes than Intel, though). Making estimates about the logic portion of the chip is harder; 1 pipeline would suggest roughly 1/4 of a Geforce256, or ~5 million transistors (although this is obviously just guessing, assuming that the internal architecture is not too dissimilar), which should correspond to about 4 mm2 in 90nm; the additional video/audio/other features will also add an unknown amount of logic.
 
Back
Top