"Claus Reinke"
either be slower than mainstream hardware or would be overtaken by it in a very short space of time.
i'd like to underline the last of these two points, and i'm impressed that you came to that conclusion as early as the eighties.
Well, Stuart and I had just been working on TIM, and Thomas Clarke was one of the folk who built SKIM, so we had a fair bit of relevant experience between us. It must have been between 1986 (when we'd finished TIM) and 1989 (when I got ill).
unlike earlier designs, the hardware only leaned toward functional, rather than being specific to it (mostly RISC, with large register files organised as very fast stack windows for a small number of stacks),
We had that sort of thing in mind -- my first implementation of TIM was on an ARM, so I knew something of what RISC had to offer.
and numbers from the hand-configured prototype suggested that it would be about twice as fast as contemporary standard hardware. which was great, until it became clear that, in the time it would have taken to go from that prototype to production, the next generation of that standard hardware would have been on the market, also twice as fast (with the next next generation already on its way)..
Yup, that's what we figured (we knew the ARM guys too, and knowing the rate at which they worked probably helped us see things quite clearly :-).
the suggestion that the mainstream might be running out of steam along one particular dimension is interesting, but in my naive view, there is still the difference between any one-shot research project and a snapshot in a development pipeline of great momentum
Yes. I think the best bet is to get hold of prototypes and research fp implementations for them; my TIM implementation of Ponder must have been the first fp language on ARM, and the process of doing that probably informed much of the detailed design of TIM. Doing an fp implementatio for a huge number of cores strikes me as an exciting avenue. -- Jón Fairbairn Jon.Fairbairn@cl.cam.ac.uk