computers are dead, long live computing

The venerable Von Neumann architecture, the predominant/only computational structure found in the modern world, pisses us off.

[nonneumann]

The first clear statement of the ‘pipeline dataflow’ was given by Conway (1963) more than twenty years ago. The use of dataflow as a programming or program-structuring technique has been developed by a number of workers, including McIlroy (1968), Kahn and MacQueen (1977) and Yourdon and Constantine (1979). In 1974 Gilles Kahn made an extremely important discovery [see Kahn (1974)], namely that the behavior of a pure pipeline functional net is exactly described by the fixed point of a corresponding set of equations. [Faustini (1972) established the validity of the “Kahn principle” for a very general operational model.] Kahn’s work stimulated a further interest in the theoretical aspect of dataflow, mainly in the problem of extending his approach to handle nets with nonfunctional nodes.

"Lucid, the Dataflow Programming Language."

[nonneumann]

Explorations in J!

JNotebook - Explorations in J, the non-von-neumann programming language by KE Iverson, influenced by APL and the work of John Backus.

[nonneumann]

[nonneumann]

Only time will tell if C is the right abstraction for the future, since it was designed for a sequential execution model rooted in Von Neumann architectures and existing design trends suggest increasingly heterogeneous and parallel architectures.

FPGA Programming for the Masses

David F. Bacon, Rodric Rabbah, Sunil Shukla, T.J. Watson Research Center

[nonneumann]

Projects to Watch

XMOS - a mashup of programmable logic and a RISC chip, chips with one and four processors are available, each offering eight hardware-level threads allowing even the fastest communication protocols (high speed USB over ULPI) to be bitbanged with other threads available for other I/O operations or calculations.

GreenArray - startup silicon designer offering the GA144 consisting of 144 independent, fully asynchronous 18b computers.

Silego - silicon designer specializing in low-cost power efficient chips, offering the GreenPak 2, a tiny mixed-signal arrays offering completely remappable pins, a handful of lookup tables for digital functions, analog comparators/digital converters, and a SPI slave peripheral. The GreenPak 2 offers a fully graphical development environment

OpenCores - open project working to implement an ASIC of their OpenRISC CPU.

[nonneumann]

 This slideshow seems to argue that computers haven't done much in the past 50 years besides make us have to think less, which is not the same as a higher quality of living.

[nonneumann]

TL;DR - It doesn't.

[nonneumann]

Discusses a variety of approaches to asynchronous processors. Same sort of deal as GreenArray.

[nonneumann]

README

I contain all the punchlines from the ALA document, with a lot less theory. ALA's valuable background, but a bit dense as a concept intro.

[nonneumann]

This is Conal's modern take on Backus's classic. It's a little Haskell oriented, but conveys the same sentiment with a lot less algebra.

Conal argues that the Von Neumann architecture, by virtue of only facilitating computation through sequential operations, is inherently imperative, and thus has serious failings with regards to both composability and IO.

Rather than simply use hacks like Monads to provide a psuedo-functional interface to inherently imperative IO calls, he suggests that we would be better off re-evaluating the fundamental assumptions of computer science, namely the C-and-VonNeumann world we live in.

[nonneumann]

GreenArray is putting 144 asynchronous microcpus on a single piece of silicon. They seem to be stuck on explicit message passing rather than implicit dataflow. They're also using imperative Forth, rather than a perhaps more elegant functional approach.

The microcpus are detailed in this design document. It's stack based, with application stored either in RAM or ROM.

[nonneumann]

For an introduction, read the paper about RALA, not this one.

I’d also like to thank my colleagues at these and other companies, especially Kurzweil Technologies, Inc., where I worked for three summers.

This is a fairly comprehensive description of an possible implementation of asynchronous cellular automata, including a simulator and implementations of bubblesort and a rudimentary encryption algorithms. The "asynchronous" part refers to the fact that operation execution is data dependency driven, rather than clock driven. The "logic automata" part is a novel implementation of a computational structure implementing a single boolean logic operation(defined by neighbors) in each cell, rather than a single dynamic bit.

Dalrymple also spends time discussing a bucket-brigade/token approach to information transfer to minimize energy lost.

[nonneumann]

Despite the fact that Backus spends half of this paper getting into algebraic proofs of why functional computing is awesome, he makes some very clear points with regards to the failings of the Von Neumann arch, coining the term "Von Neumann Bottleneck" to describe the word-wide bus between persistent memory, random memory, and computational element.

Thus programming is basically planning and detailing the enormous traffic of words through the von Neumann bottleneck, and much of that traffic concerns not significant data itself, but where to find it.