Slashdot Mirror
Paint-on Laser Brings Optical Computing Closer
holy_calamity writes "New Scientist has a story about a laser made by painting a solution of semiconductor crystals onto glass. It could be used to break the interconnect barrier by having optical interconnects, the interconnect barrier threatens Moore's law unless a faster way of connecting chips is found."
Optical interconnects could make for far more reliable connections between system components. Ribbon cables etc break easily, and are a real nightmare for assembly. OTOH, a few specks of dust in an optical connection could cause a lot of grief (reflection etc) making one wonder what the longterm prospects of shipping optically connected products are.
Engineering is the art of compromise.
it wouldn't be that hard to change it from binary (light on/light off) to amounts of light (of course the same thing could be done with electric interconnects but by voltage--just not as consistently)
Didn't the Russians try this at one point? If I remember right, they had trouble distinguishing between the on state, and the not on/not off state, and wound up abandoning the idea because it was too unreliable.
The bottleneck in computing isn't Moore's Law of transistor density. It's programming paradigms. We're wasting the vast majority of processing/memory/transmission capacity with linear programming, rather than parallel programs. Procedural programs are based entirely on the bottleneck paradigm, with the entire system reduced to a single boolean operation at any given time. Any parallelism is exceptional, and difficult to express in the symbols humans send to computers.
Parallel dataflow and distributed control are long overdue to the mainstream. Compilable UML is a slow, crude path to it. When I can draw a flowchart of primitive objects, any of which are packaged procedures or other flowed objects, and watch it run, I'll have a much better shot at exploiting all the compute/storage/transmit capacity available at that time. When "compilers" can distribute my data among the resources according to topology and analytical prediction, I'll finally get full use of the machines I'm using. Until then, I'm doubling my HW capacity every year or two so it can use half the efficiency gain running inefficient software.
--
make install -not war
For example, suppose you wanted something to operate at 10GHz. Now suppose that the medium you use is such that the wave moves at .5c. That gives you a wavelength of just 1.5cm. That means on larger dies, you can start having signal propagation problems, in that you won't get a wave all the way across the chip before the next ones starts. Plays hell on synchronized processor designs like we use today.
It's not a problem yet, that I know of, but something that we have to think about in the future.