Fusion die-shot - 2009 Analyst Day

[Acert93]

Why not an optional edram chip like xenos? How much memory would it take to handle 1920x1080 with 2xAA?

I believe just shy of 32MB using an FP10 color depth (double that for FP16).

EDIT: Not to mention some may want a more percise Zbuffer, too. And that die will get bigger if you want more flexibility in reading/writing and better bidirectional communication between dies (dice... need to check the verbiage again!) And when you consider various buffers, MRT, and possibly more texturing options ...

 

[fellix]

32MB of eDRAM would be just enough to fit a whole 1080p 8bpc back-buffer with 2xMSAA without sorting to tiling, or 720p with 4xMSAA.

 

[Acert93]

Is that a good use of 100mm^2 to 150mm^2 + of silicon? That is probably bigger and more expensive than a lot of low end GPUs!

 

[Groo The Wanderer]

The whole single chip dynamic changes a lot, and honestly it sounds like something that marketing would be able to work well with. A 3ghz quad core fusion CPU with a 480SP DX11 GPU, is definitely something with marketing potential.

Especially if the market you are in is selling CPU pins, or more importantly, selling high layer count board layout tools.

-Charlie

 

[rpg.314]

32MB of eDRAM would be just enough to fit a whole 1080p 8bpc back-buffer with 2xMSAA without sorting to tiling, or 720p with 4xMSAA.

If it has to be rendered in tiles, then is it transparent to the developer? If you don't have to worry about it in the sw, then you can obviously get away with lower edram. Having said that, I don't see edram going mainstream in gpu's anytime soon, despite it's advantages. Putting an edram inside a low end cpu package will be costly.

 

[fellix]

Sure, but not for high-volume console SKU.

Curiously, how much a POWER7-derivate core could shrink, let's say for 32nm process tech, minus the mainframe functionality -- three CPU cores, shared L3 and eight SIMD DX11+ IGP from ATi all in sub-300mm² die. 32MB eDRAM die with the all ROP guts within should be minuscule by the next platform refresh cycle. A single-chip solution, wired with 2GB 6Gb/s GDDR5.

 

[fellix]

Is that better:

6 SIMD multiprocessors, 240 ALUs, 24 TMUs, eight ROPs and full array of display outputs.
The CPU part does seem to have all the four HTT ports?!

 

[no-X]

Yes, this dieshot is better, but layout of some functional units is different

 

[rpg.314]

Probably they'll be putting up the GDDR mem controller into the chipset, ad to feed the beast they have 4 HT links.

 

[Mintmaster]

That is going to be one sick chip for mobile parts if they can get low enough power consumption during light use. I hope they can feed it enough bandwidth.

 

[DemoCoder]

Mobile as in notebook, it's way too big/power hungry for phones, and probably a little too much for netbooks. But it looks like a clear winner for medium to full size notebooks. (it's questionable whether such a chip is even needed for small screen form factors anyway)

This is definitely an exciting development and could end up saving AMD's bacon in the long run. Apple will be a tough sell even though it is ideal for some of their product line (Macbook Air for example), because of the high level partnership with Intel. I definitely wish I could get something like this in my MacBook.

 

[Raqia]

Mobile as in notebook, it's way too big/power hungry for phones, and probably a little too much for netbooks. But it looks like a clear winner for medium to full size notebooks. (it's questionable whether such a chip is even needed for small screen form factors anyway)

This is definitely an exciting development and could end up saving AMD's bacon in the long run. Apple will be a tough sell even though it is ideal for some of their product line (Macbook Air for example), because of the high level partnership with Intel. I definitely wish I could get something like this in my MacBook.

Apple didn't shed any tears over dropping PowerPC nor using nVidia's IGP chipsets instead of Intel's. They already use ATI graphics processors and dropping in some x86 compatible CPUs into the mix shouldn't entail a major change; in fact, moving to x86 itself was probably partly motivated by the fact there's more vendors than just IBM. It's not as if OSX uses any Intel only extensions or that there's a big shiny "Intel Inside" logo on macbooks that people look for when they go shopping .

 

[Crossbar]

I believe just shy of 32MB using an FP10 color depth (double that for FP16).

Technically FP16 would require about 50% more space than FP10.

 

[Mintmaster]

Mobile as in notebook, it's way too big/power hungry for phones, and probably a little too much for netbooks. But it looks like a clear winner for medium to full size notebooks. (it's questionable whether such a chip is even needed for small screen form factors anyway)

My bad, that's what I meant, although it could be great in 12-13" notebooks too (not sure if you call that medium or not). I'm envisioning a fantastic 3.5-4 pound notebook that I can use for pretty much everything.

If OpenCL apps get a traction by then it'll be pretty good for AMD as well. This is the kind of thing I was talking about before where open standards won't be able to protect the market that NVidia is trying to create.

 

[brain_stew]

Ranger I think that IGP or fusionhas still to reach the point of "good enough" in a few year IGP/fusion may deliver graphics that are good enough to please the whole range of costumers. We 're still no there tho.

Surely a 480 ALU DX11 chip is going to go a heck of a long way to solving that, no? ATI's current integrated part is a 40 ALU part, is it not? Whilst bandwidth is always going to be an issue, we've never seen a leap of that magnitude from one integrated graphics generation to the next, and then there's the tighter CPU integration which can only help performance. Heck, ATI's latest integrated graphics refresh was just an OC and die shrink of a two year old part, and such a pathetic upgrade is commonplace in that segment.

My bad, that's what I meant, although it could be great in 12-13" notebooks too (not sure if you call that medium or not). I'm envisioning a fantastic 3.5-4 pound notebook that I can use for pretty much everything.

If OpenCL apps get a traction by then it'll be pretty good for AMD as well. This is the kind of thing I was talking about before where open standards won't be able to protect the market that NVidia is trying to create.

According to the slides, the chip will be targetted at "thin & light" notebooks so the sort of machine you're envisioning is one target market AMD have already identified. Oh, and I agree a 13", sub 1 inch thick notebook with one of these chips would make an excellent machine.

 

[Kaotik]

Is that better:

http://img4.imageshack.us/img4/4966/apuu.jpg

6 SIMD multiprocessors, 240 ALUs, 24 TMUs, eight ROPs and full array of display outputs.
The CPU part does seem to have all the four HTT ports?!

Huh? That image clearly has 12, not 6 SIMD multiprocessors

 

[Blazkowicz]

Technically FP16 would require about 50% more space than FP10.

doesn't a FP10 pixel take 32bits and a FP16 one 64bits?

 

[OpenGL guy]

doesn't a FP10 pixel take 32bits and a FP16 one 64bits?

The original quote wasn't quite correct.

Why not an optional edram chip like xenos? How much memory would it take to handle 1920x1080 with 2xAA?

I believe just shy of 32MB using an FP10 color depth (double that for FP16).

FP16 doesn't double the Z buffer requirements so ~48 MB is required for 1920x1080@64bpp color and 32bpp depth buffer w/ 2xMSAA.

 

[hoom]

Huh? That image clearly has 12, not 6 SIMD multiprocessors

I see 240 unless there is still more chopped off.
See RV770 die for comparison.

 

[AnarchX]

240SPs and 24 TMUs seem really odd for a 2011 GPU-computing design?

Also the transistor density seem very low, if we consider ~100mm² for the CPU part, coming from Propus@45nm 169mm². The IGP should be the same size.

A 32nm bulk motherboard GPU with GDDR5 sideport, should be a wiser move compared to this.