| drkohler said: Coinsider one of the holy laws of electronics: P = I^2*R. This says that the power you dissipate in an element with resistivity R goes as I squared. So if you double the clock rate, you must get rid of four times the heat (since you essentially have to double the current) . In reality things are much worse because the higher the clock rate, the more leakage currents (electrons that flow to places where you don't want them to go) increase. Particularly when you shrink the process node. At some point the current goes through the roof and you instantly burn your cpu. The Jaguar cores can easily go to 2Ghz (producing roughly 60% more heat), but doubling to 3.2GHz is off limits. If you tried to increase the speed of the gpu part in the die, you'd run into thermal troubles much, much faster than with the cpu, because there are so many tiny processors in those CUs that you will soon be unable get rid of the heat fast enough. Even a "puny" 20% speed increase in the gpu part stresses the design to the limit or over it. |
That's specifically why I avoided asking an increase in GPU clocks. 
(We all know what happened to nvidia fermi, now don't we ?) There are other ways of increasing the CPU clocks such as redesigning the CPUs for a bigger die size with a longer pipeline. The bigger die size is used to mitigate the increased power per area and the longer execution pipeline will make it more scalable in terms of increasing frequency. I doubt it will be too hard to increase the clocks to 3.2 Ghz because the AMD vishera was able to do it with the same amount of cores.
| drkohler said: Doubling the functional units obviously doubles the die space those units require. The PS4 apu is approx. 320mm^2 in size. Doubling the gpu would increase the size to probably around 450mm^2. Unfortunately now you have the problem that you can't really feed all the gpu units anymore because you just can't get enough data into them. So you are more or less forced to use a wider bus, which costs you even more space for another 1-2 gddr5 controllers. These again take up so much space on your die that so you rapidly pass a die size of 500mm^2, which is generally considered an economically viable limit (also making such huge chips is an engineering nightmare). Over 500mm^2, that would be a giant chip, NVidia has those made, and they cost a fortune to manufacture. |
Getting a wider bus is extremely important and that especially goes for the AMD GCN architecture because it is known to be very bandwidth constrained. (Well there are other ways to increase bandwidth such as waiting for GDDR6 chips to hit the market because the only difference that it has from GDDR5 memory modules is the doubled input/output frequency but another thing to keep note is that new iterations of memory modules are known to be very expensive at first so I think consumers will probably also disagree with the $1000 price tag too. 
) Giant chips like the nvidia GK110 are known to be nightmares in terms of yields and their extremely expensive.
| drkohler said: In the example for Sony, and as a mental exercise, I'd go the following route: Go to bed with Samsung, cooperate with their 20nm process technology. Make 1Gbit gddr5 chips. Adapt the Jaguar (already existing in labs) successor to Samsung design rules. Increase the CU count to 24 (and some more TMUs). Add another 64bit gddr5 controller, getting ram up to 10G with 10 chips. Increase speeds depending what the 20nm process allows for, probably 10%, 20% probably is the best hope. This would give you a PS4.5 with about 250% cpu power and about 150% gpu power in a die not much bigger than now, not much hotter than now, within about 3 years. |
I was thinking of waiting for TSMCs 16nm process node or global foundries 14XM.
| drkohler said: All this can be done, there are no technical obstacles preventing that path. Will they do it? Very unlikely, but one can always dream. And in three years, other design philosophies are "more hip", or console gaming is dying fast.. |
This part is very much true and I agree.
