In a recent blog post, Sun's Santhosh D'Souza describes a new technique that allows a large number of processors to communicate wirelessly, yielding much higher bandwidth for parallel systems than wired communications between chips would allow.
While only tangentially related to enterprise software, Sun's research lab has been working on an interesting technology that, in time, can lead to yet another type of parallel system: one where individual processors communicate wirelessly. In a blog post, Proximity Communication: Coming to a Chip Close By, Sun's Santhosh D'Souza describes the new technique:
Proximity Communication lines up a transmitter on one chip against a receiver on another chip, with the two chips positioned very close to each other. The two form a capacitor and a voltage driven through the transmitter will result in a corresponding charge on the receiver, thus achieving communication. The method has been shown to achieve an I/O pad density 60 times greater than that possible with [communication occurring through wires]. It thus allows larger bandwidths for a given area on the chip.
As the chips are not wired up to each other, the technique also permits chips to be easily replaced...
The advent of multi-core processors has generated renewed interest in concurrent programming. Systems based on proximity communication can grow to possibly much larger sizes than parallel systems that must fit in the confines of a single CPU. Yet, proximity communications also introduces the network as a component of a parallel system.
Discussions about concurrent programming for multi-core CPUs can conveniently ignore the possibility of communication failure between individual CPU cores. That will not be the case with proximity communications-based systems that, in one sense, will act as multi-core processors—given the high bandwidth between the processors—and, in another sense, as distributed systems susceptible to network failures.
What do you think of proximity communication as a mechanism for building parallel systems?
Proximal wireless communication looks like a good way by which to chips can communicate with one another. If each chip has two such ports, then we could link chips together in some sort of list or tree structure.
A tree structure sounds more promising than a list structure, so that communication time between two arbitrary chips would be proportional to log(n) where n is the number of chips.
This sounds like a good technology for implementing the Mago Machine parallel architecture that was investigated during the early 1980s at the University of North Carolina at Chapel Hill.