Previous | Next --- Slide 18 of 50
Back to Lecture Thumbnails
RICEric22

Is the image not supposed to be a grid, but rather the connection between for example 0 on the left and 3 on right goes 'under' the nodes below 0 through 2 and is only actually connected at the node under 3?

cwswanso

In my understanding, no. It would still have to go through the switches along the way. Though I suppose you could implement it that way, there would just be a ton more links in the network.

ycp

I agree with @cwswanso. If 0 on the left had a separate wire going to every endpoint on the top, sure there would be a great deal of speed up, because there is a dedicated link between endpoints, but there would also be a huge problem for scalability. Then, adding another node on the top would require that each node on the left be capable of having another wire attached to it. This could also get confusing as an endpoint could start receiving information on multiple lines.

benchoi

With a separate wire from every node to every other node, there would still only be N^2 wires, which doesn't seem worse than N^2 switches (shouldn't a wire be simpler than a switch?).

I'm curious to know, though: does a wire necessarily consume power all the time the chip is running, or only when data is being sent through it? If the latter is the case, would it be right to say that whenever a crossbar topology can be used feasibly, it is always desirable to do so (barring cost?)

nrchu

I think there are physical limitations on the number of wires you can have emanating from a single node. Having N wires attached to a single node doesn't sound scalable at all.

woojoonk

So for this case any combination of nodes connecting together does not really influence other combination from communicating with each other. Since it is unlikely that there is a case where all nodes need to reach all others, it seems that blocking/non-blocking issue may not be too much of the importance when network is well structured?