[ericjang]

I wasn't aware that Danny Hillis was a student of Minsky's and that the Connection Machine mainframe series was created with brain simulation in mind. That's really cool.

When thinking about implementing the network topology for a large-scale brain simulation, the network topology should reflect the 3D spatial local-ness of the real brain (to avoid redundant N x N communication between units). One seems to either arrive at a fat-tree CM5 architecture or a 3D lattice of asynchronous processors (but this is not very general-purpose).

https://www.youtube.com/watch?v=blvC0DA96dI

[lispm]

Not necessarily brain simulations, but generally connectionist AI software. Put a graph into the machine and have operations on the graph be computed in a parallel fashion. Propagate information over the graph connections...

Thinking Machines as a company struggled with defining the target market: was it a machine for AI algorithms or was it a more general machine. See it's use with Fortran, doing more traditional stuff, not AI.

[versteegen]

But a 3D lattice is actually quite bad in the sense of having a very high graph diameter for the same degree (6) as the number of nodes goes to infinity, compared to other designs, though probably not so bad for low n. Still far better than a square lattice, of course. Wiring of processors in clusters and supercomputers is one of the main uses of graphs with low degree and high diameter (see [1]). I also know that twisted hypercubes were used for connection graph of nodes in supercomputers, because they have a number of nodes which is a power of 2, fairly simple and uniform routing rules (in general other graphs don't), large bisection bandwidth, low diameter, high (maximal?) fault tolerance, and more.

[1] https://en.wikipedia.org/wiki/Table_of_the_largest_known_gra...

[mark_l_watson]

I was fortunate enough to be able to write software for the first version of the CM (the SIMD one). Although I was keenly interested in AI even back then, I viewed the CM as a great system for parallel processing, with good I/O throughput, and not for AI research.