Reversible computing & MBrains [was Re: Matrioshka Brain detail]

Robert J. Bradbury (
Thu, 2 Dec 1999 09:57:20 -0800 (PST)

Early (very early) this morning I wrote about computational limits:

> These are the Bremermann & Bekenstein bounds. (If you want access to the
> relevant papers, send me an off-list request.) The problem has relatively
> little to do with "resistanceless" computation and much more to do with the
> cost of erasing a bit. Interestingly enough, you can do "ultra-cheap"
> computing so long as you don't erase bits. This is what gives rise to the
> reversible computing methods (that multiple groups are working on).

I realized after some sleep that this isn't correct. The original classic paper on heat in computing, was:
R. Landauer, "Irreversibility and heat generation in the computing process," IBM Journal of Research and Development 5:183-191 (1961).

Bremermann developed some theories on quantum limits to information processing and Bekenstein worked on entropy and communications costs. Charles Bennett expanded some of Landauer's work related to the thermodynamics of computing. Later the theories were worked into some more practical applications by the MIT people (Margolus, Toffoli & Fredkin). Eric of course put in his rod-logic reversible design. JoSH (Storrs Hall) extended this into algorithmic reversibility, and Eric & Ralph extended the size scale down a bit (controlling electron movement rather than atom or molecule movement). Its at the stage now where I believe that groups at MIT and UCSD are actually working on CMOS based reversible computers (which we presumably will need as clock speeds go up and you start trying to get rid of 100+ watts per chip).

The basic thing to remember about reversible computing is that it minimizes the heat production that comes from erasing bits (documented by Landauer & Bennet et al, but does so at the price of real time (having to "undo" your calculations) and/or circuit complexity since you have to add extra logic to create the gates that save the state information that gets "erased" in current designs -- more complex circuits equals bigger chips equals increased propagation delays.

At any rate, some of the bibliographies and articles may be found under:

Now, having kicked myself sufficiently --

On Wed, 1 Dec 1999, Wei Dai wrote:

> On Wed, Dec 01, 1999 at 04:33:20PM -0800, Jeff Davis wrote:
> > Second, (and here I think I'm probably gonna put my foot in it) isn't there
> > going to be a type of "computronium" which will operate in a
> > superconductive regime? Won't that "resistanceless" condition make it
> > possible to function virtually without generation of entropy/waste heat (I
> > have heard of a fundamental principle of computing/information theory that
> > assigns a minimum entropy per op/(state change?), but I have also heard of
> > some other theory of reversible or quantum computation which suggests a
> > means to circumvent or drastically reduce this minimum entropic cost;
> > though such theories are waaay over my head.)
> I think you're on the right track. From the perspective of thermodynamics
> the Matrioshka design is highly inefficient since each computing element is
> polluting everyone else with waste heat.

No, this isn't true. The standard M-Brain architecture I designed, radiates heat only in one direction (outward, away from the star). Each layer's waste heat becomes the power source for each subsequent (further out) layer. To satisfy the laws of thermodynamics and physics, you have to get cooler and cooler but require more and more radiator material. At the final layer you would radiate at the cosmic microwave background (or somewhat above that if you live in a "hot" region of space due to lots of stars or hot gas).

What you say would be in an M-Brain Beowulf cluster architecture (multiple stars, multiple M-Brains). They would be polluting each other with waste heat.

See the pictures: to get an idea of the nested shells and the size scales & possible element composition of the shell layers. Each shell layer orbits at the minimal distance from the star (to reduce inter-node propagation delays) while not melting from too much heat. [That makes the best use of the computronium in the solar system since the different materials from which computers may be constructed (TiC, Al2O3, Diamond, SiC, GaAs, Si, organic, high-temp-superconductor, etc.) each has different "limits" on operating temperature.] I suspect that some layers may be element constrained (e.g. GaAs) and assume that diamondoid rod-logic computers are not "best" for every operating temperature -- single-electron Si-based computers, or high-temperature copper oxide superconducting computers may be better in specific environments.

However it is important to keep in mind that the mass of the computers in a node is probably very small compared to the mass of the radiators and cooling fluid (this is the part that needs to be worked out in detail).

It is worth noting that Eric & Ralph's Helical-Logic architecture proposed controlling electrons at very low temperatures (less than the cosmic microwave background) in a sapphire structure. That node architecture really doesn't fit into an M-Brain anywhere because of the problems that you pay in both power losses (to get << 2.7K operating temps) and communications delays (because you need really big radiators).

> The theory of reversible
> computation says you're only forced to generate waste heat when you erase
> information. There is no minimum entropy increase per operation as long as
> the operation is reversible. This suggests that the computational core
> should never erase bits but instead ship them out to erasure sites at the
> outer periphery of the civilization or around black holes.

Now this makes for a *very* interesting proposal. I had not considered the possibility of effectively dumping the entropy increase from bit erasure into black holes! Of course this may be easy to say in theory but difficult to do in practice (as is Frank & Knight's comments about using ballistic or relativistic particals for entropy removal from a computer). [You potentially get up to 9 orders of magnitude increased power density, if you can figure out how to transfer the waste heat to ballistic or relativistic particles.]

> The MB design assumes that energy collection, computation, and entropy
> management (i.e. bit erasure) are all done at the same place. It makes more
> sense to distribute these functions to specialists. The optimal design for
> the computational part is probably exactly what you suggested: a single
> supercooled sphere with no moving parts (except maybe at the nano scale).

I don't see evidence for this. If you supercool below 2.7K the energy costs are very large. Anywhere in the ranges below 50-100K your radiators are large so your communications costs are as well. If you try to put all of your computer nodes in one place you lose all of your power to pumping cooling fluid around.

Quoting from Nanosystems, pg 370:
"the ~10^12 W/m^3 power dissipation density of 1 GHz nanomechanical logic systems described here exceeds any possible means of cooling"

In other words it is impossible to construct a cubic meter of computronium (that operates for very long)!

What most people don't know (because Eric didn't state it directly) is that the velocity of his phase change coolant through 1 cm^3 nanocomputer is very high (close to the speed of sound). If you go through the fluid dynamics equations, you discover if you push the pressures up too high you go from non-turbulent flow to turbulent flow and the power requirements increase significantly. Obviously you can't increase the fluid pumping pressure beyond the strength of the materials and there is a tradeoff between more material (to have higher pumping pressures) and reduced heat conductivity (thicker material conducts heat more slowly) and greater communications delays (because the distances between the nanoCPUs is increased).

Optimizing these designs is going to be very very tricky.