I think the most game changing features are its impressives
capabilities, but its impressive *flexibility*. Even if performance
wise, relative to power and physical volume, it does no better than
other architectures, it is still a full system on a chip, with one
crucial difference: nearly all the design is pushed up at the software
level.
To me, this is the one chip to rule them all. It would make
"specialized" hardware way easier and way cheaper, and would make
software rot less of a problem (shapeshifting architectures will
probably be more durable and less numerous than crystalized ones).
Totally breaking backward compatibility would be less problematic.
Repurposing old hardware would be much easier.
And of course, we could finally optimize our hardware for our beloved
problem oriented languages, instead of using the hardware we have, which
is often optimized for C. When we have brevity *and* performance,
programmers won't have any excuse left.
Loup.
Daniel Gackle a écrit :
I watched the video and got excited too. Petabits of on-chip
non-volatile storage? that also can do logic? That's more than a
game changer.
But it seems that HP's memristor claims are controversial within the
research community:
http://vixra.org/abs/1205.0004
http://www.slideshare.net/blaisemouttet/mythical-memristor
Some of the dispute is about priority, which may not matter so much; I
care less about *whom* I get massive on-chip non-volatile storage from
than that I get it at all. But that too appears to be under dispute
(e.g. "Myth #3" in the second link above).
I would love to hear more from people who know about this.
On Tue, Jul 10, 2012 at 12:22 PM, David Barbour <[email protected]
<mailto:[email protected]>> wrote:
Thanks for bringing this to my attention, Shawn. Real memristors
could seriously change the programming landscape, and have much
potential for directly embedding dataflow programming models and
neural networks.
I think object dispatch and imperative C programs won't be the most
effective use.
On Mon, Jul 9, 2012 at 11:23 PM, Shawn Morel <[email protected]
<mailto:[email protected]>> wrote:
Just watched a very interesting talk on memristors:
https://www.youtube.com/watch?v=bKGhvKyjgLY&feature=related
I hadn't bothered going into very much detail so far - for some
reason, I thought memristors would end up being primarily used
as memory elements that supplant the traditional sram, dram, HDD
hierarchy. That on its own is kind of cool and would probably
help shift us away from files and more towards long-lived objects.
The talk, however, describes ways that memristors can be
organized to be an arbitrary combination of switching, memory,
logic or even analog emulations of synaptic behaviour. The talk
touches briefly on compiling from C down to logic gates
(Russell's material implication). Some key aspects is that, as
opposed to FPGAs the "reprogramming" can take place in a very
short time and they addressing capabilities of a HW associative
memory are quite large.
For example, it could take a few nanoseconds to create HW N-way
associative lookup - that's to say, I could on the fly configure
a piece of HW to actually represent object message dispatch!
shawn
_______________________________________________
fonc mailing list
[email protected]
http://vpri.org/mailman/listinfo/fonc
