Slashdot Mirror
Introducing Probability into Chip Design
prostoalex writes "The August issue of Intel Developer Update has an interview with Shekhar Borkar, Intel Fellow and Director of Circuit Research at Intel Corp. talking about the future of microprocessor design and what goes on inside Intel Labs. Borkar tells why we need even faster processors and how probability will make its way into future chip designs - "It's like the shift from Newtonian mechanics to quantum mechanics. We will shift from the deterministic designs of today to probabilistic and statistical designs of the future.""
What he appears to be suggesting is transistors that we acknowledge to be based in an analog world -- their state depends not only on the data you feed them, but also on the temperature they are immersed in, etc.
...back in the heady days of Concurrent Computer their top-of-the-line 3280 processor has "usual branch" instructions. The compiler could use the usual branch instructions to provide hints about the probability of the branch being taken to the processor. In a loop, for instance, you'd use a "usual branch not equal" (UBNE) instruction to send execution back to the top. This would indicate to the processor that it should preemptively invalidaate the cache and pipeline.
I'm sure many mainstream processors have this now, but it's funny to think that CCUR had this technology in the late 1980's.
Tiller's Rule: Never use a word in written form that you've only heard and never read. You will end up looking foolish.
i've been using a computer for so long that i'm half convinced that the reals are a hoax invented by physicists to make their sums easier ;-)
Well, at least you are, or rather were, in good company. When Lindemann proved the transcendence of pi, Kronecker asked:
"Of what use is your beautiful investigation of pi? Why study such problems when irrational numbers do not exist?"
There is a problem with your "proof."
.9999.. = 1, you're saying: .999... = 9
Let x be 0.99999.... x = 0.9999....
10x = 9.9999...
10x - x = 9.9999... - 0.9999... = 9
10x - x = 9 (the infinite trail of nines drop off in the subtraction)
^-- this step is artificially inflates the value of 10x, in reality this would be 8.99999999999999...1
So the question herein lies, what happens when you shift an infinitely repeating decimal 1 place to the left? Does it magically gain a new decimal place on the right? This may happen, but that violates the basic laws of normal numbers. These mechanics of an infintely repeating decimal cannot be used as the basis of a algebraic proof.
Because you're not saying
9.999... -
but 9.9999... is not 10(.99999...)
10(.99999...) has one less decimal place on the right than 9.9999... But still an infinite amount of decimal places. Just because they're both infinite doesn't mean it's not a smaller infinity.
Since physical science (and by extrapolation, engineering) is built on a "reductionist" paradigm where every problem is broken to its simplest components & solved piecemeal at that level, it makes sense for a "probabilistic" approach to chip design to happen some time. Might as well be now.
But when we operate under the reductionist model, we forget emergent properties at the system level. In developing a "learning" system -- which again, I assume to be the overarching goal -- we have to learn to deal with variation. Situations are almost never exactly the same. In the beginning of a "learning" system, things probably (pun intended) do look random. But as special cases, exceptions, subtle cues, etc. are encountered by the system & incorporated into the decision-making process, things appear to become increasingly deterministic.
So, if a "probabilistic" chip design is implemented properly, it likely will look pretty "deterministic" to the end-user, who expects certain kinds of results.
The problem now is that the hardware is "deterministic" & any attempt to create a "probabilistic" learning system has to happen in software. Right now, the limit to AI, IMO, is simply that chips aren't even in the same league with neurons. "Learning" software built on "learning" hardware ought to be a pretty powerful concept.
Of course, this may just be a way to get around the fact that manufacturing may be pushing the limits for tight tolerances & probabilistic chip design is the only out. Whatever it takes to force a paradigm shift.
"Most places a paradigms won't buy you a cup of coffee..."
"Obviously, I'm not an IBM computer any more than I'm an ashtray" (Bob Dylan)
He's not talking about non-deterministic computing. He's talking about ways to salvage the chip if one or more subcircuits don't function correctly. The article isn't very technical, but this probably alludes to having redundant circuits, possibly even taking the answer that the most redundant circuits produce.
I'm not a smart enough man to know whether or not this is feasible. Keep in mind that introducing these redundancy checks actually increases the "length" of the circuit, increasing propogation delays. If this system works at all, you can be certain that it will be very rigidly subjected to the law of diminishing returns.
-Roxton