Real cpu/gpu fusion due in 5 years?
http://www.xbitlabs.com/news/cpu/di...tion_of_AMD_Fusion_Chips_Due_in_2015_AMD.html
http://www.xbitlabs.com/news/cpu/di...tion_of_AMD_Fusion_Chips_Due_in_2015_AMD.html
Fusion die-shot - 2009 Analyst Day
- Thread starter AnarchX
- Start date
aaronspink
Veteran
Just because it is CMOS doesn't mean it is the same process. A DRAM device needs a capacitive structure in order to store the data and this requires either a trench process or a MM capacitive layer. Both require additional processing steps and have somewhat different metal/transistor optimization points than logic optimized processes. Even within logic processes there can be significant difference, hence you have TSMC offering like 4 different 40nm processes.
Are you suggesting that even though it has logic structures (ROPs) on it, the die is probably made with a process more optimized for making DRAM's instead of logic?
EDIT: no sarcasm intended, just an honest question.
Last edited by a moderator:
btw original Xbox 360 had 90nm CPU, 90nm GPU and 90nm EDRAM. Current 45nm CPU+GPU die measures only 47% of the original size. The EDRAM at 90nm had about 80mm² and now it has 56% (45mm²). Is it manufactured by 45nm process? It shrunk significantly less, than the CPU/GPU. Either the fusion of CPU/GPU die saved really significant area, or the EDRAM isn't manufactured using 45nm process...
Jawed
Legend
The FSB (physical and supporting logic) doesn't take too much area, but in effect 2x that area is saved, once for each of Xenon and Xenos (EDIT: hmm, no, prolly only the physical is saved, I expect logic is still required):
Meanwhile on the EDRAM die, the HSIO physical will not have shrunk much, if at all.
Jawed
Meanwhile on the EDRAM die, the HSIO physical will not have shrunk much, if at all.
Jawed
I guess the pads on edram are limiting the shrink. I/O's scale down very poorly compared to logic. Cell 90nm -> 45nm shows similar behaviour, as the floorplan wasn't modified.
aaronspink
Veteran
Yes. ROPs all told are fairly simple. Load, compare/add, write. There is logic on bog standard commodity drams to do things like mux data and run the self refresh, etc. The logic isn't that fast because it doesn't need to be. Its perfectly reasonable to go with a DRAMish process for the embedded memory. The ROPs aren't that complicated and don't need to run that fast really.
Expanding on the complexity comment... One of the advantages of edram is it allows for simpler ROPs. Compression logic adds complexity to the ROPs or Render Backends in Xenos terminology. The abundant bandwidth of Xenos' edram meant depth and color compression wasn't necessary.
Ontario has been brought forward. Amd mentioned that it would be shipping for revenue in late Q4 2010 so expect it to appear in Q1 2011. Bulldozer looks set for a Q3 2011 release and Llano has been delayed by a few months and now looks set for a late Q1 2011 release. Dirk Meyer said its because the yield on 32nm isnt high enough. Intel's been shipping 32nm processors since December 09, looks like GF is way behind. And judging by the problems with 32nm, im guessing we probably wont see 28nm products from GF until Q3 2011.
Source http://www.fudzilla.com/processors/processors/processors/amd-tapes-out-its-bulldozer-core
Source http://www.fudzilla.com/processors/processors/processors/amd-tapes-out-its-bulldozer-core
hoom
Veteran
Various sites have been reporting that the GPU section of Llano is Redwood ie 400SPs.
Currently 5670, 5570 & 5550 (cut down to 320) cards use Redwood.
Or in other words, around the same range/probably better than a 3870.
Not exactly amazing GPU power but to have that sort of thing as a min-spec for budget PCs should have good implications for the future of gaming & 3D accelerated apps in general.
Currently 5670, 5570 & 5550 (cut down to 320) cards use Redwood.
Or in other words, around the same range/probably better than a 3870.
Not exactly amazing GPU power but to have that sort of thing as a min-spec for budget PCs should have good implications for the future of gaming & 3D accelerated apps in general.
As has been commented often, bw available to it will determine it's perf levels much more strongly than the no. of ALU's it has. However, it is good to hear that AMD is also being aggressive wrt GPU perf in it's products. 400 ALU's would mean substantial amount of area has been devoted to gpu.
Last edited by a moderator: