Originally Posted by sdlvx
The problem with flagship APU that is 4m/8c with iGPU is not TDP, it's die size.
Vishera die is 315mm^2
290x is 435mm^2, so 4m/8c + 290x iGPU would be 750mm^2
Tahiti is about 320mm^2, so 4m/8c + Tahiti iGPU would be 635mm^2512GCN core APU parts are 100w peak IIRC.
Even if you assume 135w for FX 8350 level CPU + 100% of TDP of the APU going to the GPU, then that's still just a little more than an FX 9590 or other heavily overclocked CPUs.
I wrote this before I saw where this thread was heading. Imagine a system with 4 HTX slots or some sort of Freedom Fabric slots and each slot has a board with 4m/8c CPU + big GPU and they all access a giant pool of shared memory.
Yes I know latency but PCIe has more latency than Hypertransport and no one is complaining about PCIe.
EDIT: I forgot to add, I kind of viewed FX 9590 as AMD just getting motherboard manufacturers ready for 200w+ stock CPUs and APUs. AMD got motherboard manufacturers to revise their lineups to support those high end CPUs with massive TDP, do you really think they would have done that if the only product that would make that much heat was FX 9590?
Not quite. 7750s with their memory controlers, RAM and so on run straight off the Motherboard's PCI-e power of 75w. And that's on GCN 1.0. GCN 1.1 reduced power consumption by a whopping 35-40%, meaning those 512 shaders are responsible for only 45w (or less) of power.
You can get a A10-6700, a 3.7 Ghz 2-module (sound familiar?) with 384 VLIW4 shaders for 65w of power. The APU's nearest cousin, the 6570 (a low-clocked 480 sahder part) also sits around 45w, meaning the GPU's half is probably sitting at 35w, leaving 30w for the CPU.
If you include the node shrink to the CPU half as well as any further power improvements made in Steamroller, there is no reason we can't see a locked Kaveri chip at full CPU and GPU power in the same 65w envelope both Trinity and Richland enjoyed, while AMD leave it at 100w for the unlocked parts.
Also Area doesn't work that way... 435mm^2 + 315mm^2 = 537mm^2, not 750mm^2.
sqrt [(435 x 435) + (315 x 315)]
sqrt (189225 + 99225)
You'd have to cram a ton more memory controllers on there to even think of feeding it, but ya.
Originally Posted by AlphaC
Originally Posted by KyadCK
No, the 7750 is the 512 shader part. 7770 and 7790 have 640 and 896 respectively, and there is going to be a huge RAM bottleneck as it is, meaning Kaveri is already going to have a hard time keeping up with a normal (GDDR5) 7750.
7750 is about 800GFlops at 800Mhz , 512 shader.
Kaveri is 856Gflops @ 720Mhz, 512 shader
If Kaveri is being used for GPGPU (not gaming) and can hit 900Mhz to 1GHz, it'll easily be over HD7770 (~1200Gflop) , no?
If it's anything like folding/BOINC the memory clock doesn't matter as much.
AMD got about 20% more GFlops/W with Hawaii so it wouldn't be surprising ; HD7790 was a step up from HD7770 using the same type of 128-bit memory structure (doubling ACEs ala Tahiti --> Hawaii). Kaveri is also 128-bit if it's DDR3 dual channel.
The 7770 is already at 1Ghz with 640 shaders, so 1Ghz with 512 will not top it. The Kaveri GPU would have to be at 1250 to match a straight up 7770 at stock, and 1375 to match the 1.1Ghz model. Luckily there should be plenty of thermal headroom for people to enjoy.
Also the 7770 and 7790 only come in GDDR5 variants, the strongest GPU to use DDR3 is the 7750, which makes sense.
I don't think a massive overclock is completely unreasonable, but catching the 1280-1400 GFLOPs of a 7770 won't be easy, that's a 50% overclock.
Performance per watt is awesome, but AMD did not gain performance per core at all. Not this generation anyway. They just took their new wattage and threw more shaders at the problem.Edited by KyadCK - 11/12/13 at 2:06pm