Connection Machine. A missed opportunity for AI?
In many ways Moore’s Law ended at least a decade ago; the law that postulates that the number of transistors in a chip will double every two years. Early in the 21st century, the CPU speed reached around 3 Gigahertz, but since then we have seen very little progress as quantum tunnelling effects come into play when transistors reach a width of less than 5nm. It is expected that by 2026 the miniaturisation of MOSFET will end and the Von Neumann architecture may increasingly become obsolete subject that a suitable replacement is found or there is a better mitigation of its store-compute bus bottleneck. Modern AI is increasingly thirsty for more power, evident by the increasingly fragmented designs (FPGAs, GPU). Some new designs appear to take advantage of the tunnelling effects such as the switching depends on the likelihood of the electrons crossing the barrier, though a number of limitations are well documented. But most importantly, the exponential rate of change of R&D cost for smaller size transistors makes the semiconductor market an oligopolistic market.
If it was not for the politics of Cray Computer Corporation to prohibit competitor’s Danny Hillis connection machine to be sold internationally, the history of AI could have had a different realisation path. The AI winter — the long period of subdued interest in AI because of Minsky’s criticism on Rosenblatt perceptron– would have ended earlier and the Connection Machine could have had the market cap of Intel Corporation. Von Neumann’s architecture may have inadvertently slowed down the digital revolution.
We may end up witnessing a singularity event in this century, but unlikely it will happen in this decade.