←back to thread

26 points gidellav | 1 comments | | HN request time: 0s | source
Show context
Animats ◴[] No.45784884[source]
The article mention EAI's SIMSTAR, a real software reconfigurable analog computer from the 1980s. Here's a description of how that worked.[1] This was pretty good. They had a crosspoint of FET analog switches to pass signals around, so they managed to do this in solid state, without relays. They had a M68000 as a control machine to set up connections. It all worked well enough to allow a real time man-in-the-loop simulation of the F-16 control system.[2] Just barely. The USAF paper says that the biggest problem was keeping the SIMSTAR analog computers alive.

Analog FPGAs have been made, downsizing this sort of thing to chip size. But not recently. The problems seem to be 1) lack of a use case, and 2) noise.

Analog computing is all about noise minimization. This is Not Fun.

There's interest in this stuff for neural nets, which do a lot of clipping and may be less noise-sensitive.

[1] https://www.analogmuseum.org/library/simstar_technology.pdf

[2] https://apps.dtic.mil/sti/tr/pdf/ADA189675.pdf

replies(2): >>45785512 #>>45798389 #
1. Marshferm ◴[] No.45798389[source]
One has to think in different terms in signal to utilize analog. Think collisions in waves that cohere. We can’t have an end goal in hardware dev unless we have an existing software that crudely maps to the goal. We use such imprecise ideas in hardware to compute, they’re arbitrary. Our computers are still at the toy stages. Neural nets are toy models. Think organ stage.