Originally posted by: RaynorWolfcastle
Originally posted by: Vee
...Once software performance starts to scale close to geometrically with number of cores, lots of things that used to work against us, dramatic growth of power/heat, number of transistors for diminishing performance returns, and diminishing yield with die size, - suddenly starts to work for us instead. Small sacrifices result in a rich harvest of cores.
This only works to a certain extent and even then only on certain datatypes. That is, if you're proccessing video/images, this really involves applying the same operations on large datasers and lends itself well to parallel processing. On the other hand, something like AI needs fast linear processing since it's branchy by its nature.
As for optical computing, I think the main problem is that there is no such thing as cheap "optical memory" that can be generated with the kind of ease you have with electronics.
"
This only works to a certain extent and even then only on certain datatypes."
I really have to disagree with the gist of that statement. I have a hard time to think of anything performance critical that cannot be distributed on multiple threads.
"Branchy" or not has not much to do with this. You can afford to be much more optimistic about multicores.
As for optical computing, it's way hyped by media folks who don't understand the details, "speed of light" and all that drivel. Switching speed of light logic can conceivably be utterly awesome. So for that reason, some custom logic for very specific purposes could achieve insane performance. Like filters and other communication processing. But for a general CPU, I believe current transistor logic, consumer CPUs are already beyond what is even theoretically possible with optical logic.
Physical size and lightspeed will mean that signalling will permeate through light logic slower than in silicon, despite the slower switching speed of transistors.
And you're quite right about memory.
I suppose I could be wrong of course. And that there could come some conceptual logic architecture breakthrough, that means the switching speed of light could be harvested for general computing. I don't really believe that though. And even if so, I would think it would be a dead end, eventually. The future lies in being smaller than light can ever be.
We are already. And we have already made even molecular semiconductors and connected and tested them successfully. The challenge is building large circuitry with something else than current litographic integration. It's still a smaller challenge than an optical CPU, I believe.
More immediately, the key to CPU performance lies in the memory interface.
There are some possible ways to achieve a breakthrough here. Connecting the CPU directly to hundreds of MB ram, through hundreds of thousands of connections, by placing a ram die directly on top (or under) the CPU die. "Level" 4 cache?