Originally Posted by os2wiz
Originally Posted by sdlvx
os2 did you read my posts? AMD can not replace a big CPU with a big dGPU like Tahiti with any sort of APU. The TDP would be like 400w for a single chip.
I also explained how 6m/12c Piledriver on 32nm with HDL would only be 15mm^2 larger than FX 8350 dies. I think it makes sense for AMD to do this, specially since APU GPGPU is only good for floats and not integers, and Bulldozer modular design is very good for integer.
I do not think AMD can actually replace a high end CPU with high end GPU with an APU.
I think you have a big misconception. The apu is not for high end graphics. It is for mainstream graphics. You want highend graphics you add a discreet card. A six or preferably 8 core excavator apu is totally feasible for size of die on 22nm process while tdp could be about 85 watts. Where you got the notion that the graphics core in the apu was to be high end, I do not know. It also aids in HSA.
No, the point of APU is to do GPGPU workloads (physics calculations, rendering, video transcoding, etc) that are heavily threaded. The decent gaming performance in a small package is just to get hardware adoption up.
I think you are missing the real point of APUs. Yes they are great chips for gaming right now if you're on a budget but that's not the end goal. iGPUs have been around for a while and they aren't special, AMD's special sauce is going to be dumping GPGPU workloads seamlessly into the GPU whereas they would normally be done on the CPU. It is much like the idea of Xeon Phi, and I would argue that Phi is a copy-cat of HSA, but it is early and not done as well.
If you look at HSAIL there is nothing about APUs or whatever, it's all about getting CPUs and GPUs to work together better. There is nothing in these documents that state that you can't
have a system with 4 16 core Opterons in it and 4 dual GPU, big GPU cards in there. Right now PCIe is kind of a problem but there are ways around it and AMD has patents which would let them get around the issue of not enough bandwidth just fine (HTX and PCIe hybrid slots).
Hypertransport 3.1 already has more bandwidth than PCIe 3.0x16 slot and HT 3.1 came out in 2008. Hypertransport 4.0 would smoke PCIe3.0 x16, probably even PCIe 4.0. It is what is needed and it would probably warrant a platform upgrade, which may be why AMD has postponed or cancelled FX Steamroller.