Bah, I've got more.
Bubble sort was invented because it's optimal on a Turing machine, and it's easy to laugh at (ha, ha, topology matters after all). Of course, Turing chose the most reductive topology (linear tape) to simplify proof mechanics, and not as a realistic topology for any computation, ever.
Two fundamental technologies define our current electronic regime:
- Planar process — 1959
- CMOS — 1963
The first puts us into a fundamentally 2D electronic topology at the lowest scale. The second determines the first-order term in thermal efficiency. We've been running these rapids for my entire life—and for the better part of 50 years, it never blinked.
At the layer of the data center, with all those high-speed cross-bar switches pancaking the fabric, the meta topology is closer to 4D at the logical level (switching latency), and 2.5D at the physical level (speed of light effects). But still 2D at the silicon level. (It's only recently that TSV HBM is starting to appear in GPUs targeted at neural networks. Call that 2.5D.) With electronic switching, the 4D term presently dominates, but with the advent of photonic switching, the 2.5D term will likely dominate (alongside a one or two order-of-magnitude improvement in data-center bisection bandwidth).
When you look at computation on a planetary scale, and we're back to 2D (so far we mostly install our computational devices in the razor-thin planetary biosphere).
Now the human brain is 3D volumetrically, but probably closer to 2.5D at the logical connectivity layer. Back up to 3D at the level of individual neural networks. (Is that important?)
The band seems to range from 1–70 Hz. This is not dissimilar to planetary Internet-scale resonant frequencies: light circles the equator at about 7 Hz.
Human social intelligence resonates on the scale of seconds to minutes (your average drunk can thumb-select a wry emoticon for his Twitter feed in about the same length of time it takes to eject a floppy disk and jam in a different one—also known as 10 billion clock cycles). Machine social intelligence—should this come to pass—will resonate on a scale somewhere in the milliseconds to low seconds range.
The cleavage points in the time domain are strikingly different, yet more or less the same cup of tea, all the same.
This argument from time is hardly decided. An argument from Joules would probably be more useful, but is presently hard to assess in any shallow way.
What's the asymptotic data-recall efficiency of photonic memory?
Right. I've got no clue, either.