james_nicoll: Other people must have looked at this

james_nicoll (

james_nicoll) wrote,
@ 2009-07-09 14:36:00

Other people must have looked at this
Once Moore's Law no longer holds due to inevitable natural limits, what are the logical strategies the computer industry will follow?

(59 comments) - (Post a new comment)

	twoeleven 2009-07-09 06:38 pm UTC (link)
	Increasing use of parallel processing, which we're already seeing. (Reply to this) (Thread)

	montoya 2009-07-09 06:40 pm UTC (link)
	That still requires Moore to be operable, because it requires increased transistors per dollar. (Also, I'd argue that cheapification is likely to be a bigger trend than parallelization. A 16-core processor is not twice as fast as an 8-core one, but it is twice as expensive.) (Reply to this) (Parent)(Thread)(Expand)

(no subject) -

twoeleven, 2009-07-09 06:48 pm UTC (Expand)

(no subject) -

keithmm, 2009-07-10 04:13 am UTC (Expand)

(no subject) -

brooksmoses, 2009-07-10 08:57 pm UTC (Expand)

	montoya 2009-07-09 06:38 pm UTC (link)
	Settling into being a commodity business. (Reply to this)

	jeffr23 2009-07-09 06:49 pm UTC (link)
	Software improvements, probably at the compiler level mostly, may give a few more doublings of effective power. Spending time and effort writing a really, really clever compiler doesn't have a big payoff when you can get the same speed improvements just by waiting a few months, but once that doesn't hold... (Reply to this) (Thread)

alexx_kay
2009-07-09 07:31 pm UTC (link)

I'm surprised that you're the only one to mention this so far, as it was my first thought. Software at all levels (not just compilers) is written for efficiency of production these days, not efficiency of execution. It's a rare program that couldn't accomplish the same functionality with half the resources *and* twice the speed, if those goals became important. For Microsoft programs (including Windows), I'd make that estimate closer to a factor of 10.

(Reply to this) (Parent)

	krin_o_o_ 2009-07-09 07:07 pm UTC (link)
	> what are the logical strategies the computer industry will follow? Pretending that Moore's Law is still in effect. (Reply to this)

	ckd 2009-07-09 07:07 pm UTC (link)
	Chips powered by lunar He3! (Reply to this)

xiphias
2009-07-09 07:09 pm UTC (link)

Working on battery technology, maybe? Seems to me that, once you've got things working at close to their theoretical limit, you're going to have enough computing power there for pretty much whatever most general users are going to want to do.

So then you can start working on the other parts of the computer. Screen technology is already improving by leaps and bounds, so maybe that, battery work, and maybe something with input technology?

(Reply to this) (Thread)

	fridgepunk 2009-07-09 08:19 pm UTC (link)
	super-Batteries and compact power sources will come about before moore's law reaches a wall - I'm actually surprised that there isn't more wellstone-esque tech in most modern sci-fi, battery tech is receiving so much research by the phone and car and power companies and computer industries that something impressive is likely to occur in the near future. (Reply to this) (Parent)

dd_b
2009-07-09 07:09 pm UTC (link)

Off the top of my head, I'd say:

denial
anger
bargaining
depression

Not so sure about acceptance, though.

(Reply to this) (Thread)

	daev 2009-07-09 08:36 pm UTC (link)
	This gets my vote for best response ever. (Reply to this) (Parent)

anzhalyumitethe
2009-07-09 07:25 pm UTC (link)

We'd have to stop being brute forcing everything and start being clever again.

I'd guess some attempts at new architectures.

We're getting into the ubercored systems in the HPC arena. There's talk of 1024 cored chips for processing. If you went with a superspeculative architecture along with that, you could do a faster faux single processor.

We're also talking about 1 million + processor systems now too. Which some of us think is really, really ridiculous: MTBA will simply kill it. You'd almost be precluded from EVER doing a full system job.

http://www.nersc.gov/nusers/resources/franklin/

We call it Cranky Franky for a reason.

(Reply to this) (Thread)

j_larson
2009-07-09 07:35 pm UTC (link)

I also expect serious re-engineering of existing software. There's an awful lot of code around that was optimized for speed-to-market rather than performance (or reliability, or security, but I digress...). Getting 10% better performance for a year's effort isn't attractive now, since next year's machines will be 40% faster anyway, but if hardware performance is stagnant, 10% starts to look pretty good.

(Reply to this) (Parent)(Thread)(Expand)

(no subject) -

anzhalyumitethe, 2009-07-09 07:53 pm UTC (Expand)

	daev 2009-07-09 07:42 pm UTC (link)
	I used to write assembly-language code for the Connection Machine, which is the sort of system you are talking about. It was hard. (Reply to this) (Parent)(Thread)(Expand)

(no subject) -

anzhalyumitethe, 2009-07-09 07:52 pm UTC (Expand)

(no subject) -

armb, 2009-07-09 08:31 pm UTC (Expand)

(no subject) -

anzhalyumitethe, 2009-07-09 09:42 pm UTC (Expand)

	andrewducker 2009-07-09 07:25 pm UTC (link)
	Being transhuman. (Reply to this)

autopope
2009-07-09 07:27 pm UTC (link)

BTDT and speculated about it.

The CE industry is inherently deflationary -- Moore's law conceals this because we double the number of transistors on a die each generation, but under the hood the prices are falling by c. 20% per annum. Once we stop being able to have more transistors, existing fab lines will be amortized and the products will be commoditized. I speculate that we'll then enter a period where the computer industry splits between (a) high-end well-designed premium kit (cf. Apple) and (b) cheapCheapCHEAP!!! (cf. the netbook sector). And then there'll be a huge recession and layoffs, just as there was in aerospace around 1970 when the industry hit a performance wall (note that airliners today fly no faster than they did in 1970 -- Concorde's champagne quaffing elite aside, travel at over Mach 0.9 is not commercially sustainable).

Ultimately the field will be commoditized and after a period of consolidation and mergers it will become as thoroughly boring to outsiders as locomotive or airliner manufacturing.

The interesting developments will then take place in the areas of networking and software (and even in the latter, we've pretty much standardized around some local minima as our key platforms -- see also this paper by Rob Pike in dog-in-the-manger mode).

(Reply to this) (Thread)(Expand)

	happyinmotion 2009-07-09 07:57 pm UTC (link)
	Can I be really fanboy here and provide the relevant link? (Reply to this) (Parent)

(no subject) -

anzhalyumitethe, 2009-07-09 08:27 pm UTC (Expand)

(no subject) -

autopope, 2009-07-10 01:58 pm UTC (Expand)

(no subject) -

anzhalyumitethe, 2009-07-10 05:34 pm UTC (Expand)

(no subject) -

yorksranter.wordpress.com, 2009-07-10 11:14 am UTC (Expand)

	seawasp 2009-07-09 07:35 pm UTC (link)
	Fund new-physics work and reboot Moore's Law for another millennium. Obviously. How else will we all ascend to computational godhood? (Reply to this) (Thread)

	j_larson 2009-07-09 07:49 pm UTC (link)
	And if the gods don't want company? (Reply to this) (Parent)(Thread)(Expand)

(no subject) -

seawasp, 2009-07-09 07:53 pm UTC (Expand)

(no subject) -

zxhrue, 2009-07-16 10:27 am UTC (Expand)

	carloshasanax 2009-07-09 08:00 pm UTC (link)
	Remember the Textile Singularity? A rather boring industry with some high fashion to keep the market perking, and hopefully with less coerced labor. Thus, not very different from at present, but without the messianic nerd types wandering around. (Reply to this) (Thread)

	anzhalyumitethe 2009-07-09 08:27 pm UTC (link)
	I still want better groupies. ;) (Reply to this) (Parent)(Thread)(Expand)

(no subject) -

carloshasanax, 2009-07-09 09:15 pm UTC (Expand)

(no subject) -

autopope, 2009-07-10 02:02 pm UTC (Expand)

(no subject) -

le_trombone, 2009-07-10 03:42 pm UTC (Expand)

(no subject) -

autopope, 2009-07-11 07:16 pm UTC (Expand)

(no subject) -

anzhalyumitethe, 2009-07-10 05:31 pm UTC (Expand)

	jaylake 2009-07-09 08:14 pm UTC (link)
	Heat dissipation and its Siamese twin power management. That's the core problem of moving onto the Kardashev scale, both at a macro level (climate, etc.) and a microlevel (ie, processors). Edited at 2009-07-09 08:14 pm UTC (Reply to this)

dubiousprospects.blogspot.com
2009-07-09 09:18 pm UTC (link)

There are three ways currently in use of getting a problem split across processing units -- distributed processing, symmetric multi-processing, and massively parallel processing.

Distributed processing is like Folding@Home, SETI@Home, etc.; chop the problem into chunks, pass them out, get the answers back, integrate them in the big answer pile. Works nicely, but only a small subset of problems are suitable.

Symmetric multi-processing (what happens with 2 or more cores in a processor, or 2 or more processors per machine) is slightly sub-linear in performance increase even in the best case, but is much better as a way to take advantage of improved manufacturing than trying to make one core that fast. It's mostly understood for algorithm purposes and mostly not used; there aren't that many programmers who know how to do it right, and most software doesn't take anything like full advantage.

Massively parallel is what happens in a graphics card; there isn't anything symmetric, it's stuff getting thrown into a large group of sausage machine that communicate with each other but not in a time-synchronized way. Potentially much, much faster than SMP but not understood at all for algorithmic purposes; still in the empiricism and blood sacrifice stages.

My expectation is that logical strategies have three parts, all of which we are already seeing:

Move the crunch-work into central nodes accesses via some ubiquitous communications technology.

Push the power requirements and heat budget of the device people interact with as low as possible; this makes it (more) portable, lighter, cheaper, etc., and if some or most of the crunch is happening at a central location, the experience is as good or better than having a powerful dedicated processor in your personal machine.

Work hard on improving algorithms so we can take full advantage of what we've already got in the way of symmetric and massively parallel multiprocessing.

There's also a bunch of alternative fabrication technologies out there; we may yet see iron-on-diamond or similar.

(Reply to this) (Thread)

anzhalyumitethe
2009-07-09 09:41 pm UTC (link)

Symmetric multi-processing (what happens with 2 or more cores in a processor, or 2 or more processors per machine)

Doesn't have to be. We have an embarassingly parallel machine that uses each core effectively as a separate CPU for a separate job. Our Cray uses them like independent CPUs also.

With a superspeculative architecture, you could possibly make them act like one really fast, albeit slightly unpredictable, CPU. It could be as slow as a single core, but possibly as fast some percent times the x number of cores faster. Your algorithms become very, very probability based then though.

(Reply to this) (Parent)(Thread)(Expand)

(no subject) -

dubiousprospects.blogspot.com, 2009-07-09 09:58 pm UTC (Expand)

(no subject) -

anzhalyumitethe, 2009-07-09 10:12 pm UTC (Expand)

(no subject) -

carloshasanax, 2009-07-09 10:19 pm UTC (Expand)

(no subject) -

anzhalyumitethe, 2009-07-09 11:58 pm UTC (Expand)

(no subject) -

dubiousprospects.blogspot.com, 2009-07-10 02:40 am UTC (Expand)

(no subject) -

carloshasanax, 2009-07-10 09:09 am UTC (Expand)

(no subject) -

dubiousprospects.blogspot.com, 2009-07-10 12:06 pm UTC (Expand)

(no subject) -

anzhalyumitethe, 2009-07-10 07:42 pm UTC (Expand)

(no subject) -

carloshasanax, 2009-07-10 08:36 pm UTC (Expand)

(no subject) -

anzhalyumitethe, 2009-07-10 09:23 pm UTC (Expand)

(no subject) -

dubiousprospects.blogspot.com, 2009-07-10 12:22 pm UTC (Expand)

	doug_palmer 2009-07-09 11:59 pm UTC (link)
	Asynchronous circuitry. This tends to be inherently more efficient than synchronous (clocked) circuitry, just also inherently vile to design. By the time Moore's Law runs out, computer-aided designs may be possible. (Reply to this)

	This may not have to do very much with the computer industry... amberite 2009-07-10 12:13 am UTC (link)
	...but rather its scavenger outgrowths (which affects me rather more since I've never been able to afford a computer new): http://wiki.freegeek.org/index.php/Speculations_on_the_Future_of_Free_Geek (In case you're wondering, FreeGeek is this, and it's really awesome.) (Reply to this) (Thread)

	Re: This may not have to do very much with the computer industry... carloshasanax 2009-07-10 01:19 am UTC (link)
	My guess: market segmentation of student computers (high school, college) will be enough to keep it afloat. A new computer for a new stage in life. Re-engineered game platforms as well. (Reply to this) (Parent)

	mmcirvin 2009-07-10 04:29 am UTC (link)
	Embedded system programmers take over the world. (Reply to this)

	beamjockey 2009-07-10 12:40 pm UTC (link)
	I don't know. I'm too worried about the Livingston Curve pooping out. (Reply to this)

acpizza
2009-07-14 05:53 am UTC (link)

They could always start using all the power they've put into the hardware getting to that point, and stop making software ever more bloated.

The typical OS and application suite contains far more data than the entire human genome. Ponder that for a while and the vast inefficiency that has resulted from software developers slacking off more and more and more every year will make itself abundantly clear.

Hell, these days the companies making the OSes are keenly aware of the fact that if they make it more bloated and slower, people will eventually resort to buying newer, faster computers - typically bundled with more OS licenses of course. The death of Moore's law would seem one way to put an end to this.

(Reply to this)

(59 comments) - (Post a new comment)

Username:		Create an Account Forgot your login or password? Login w/ OpenID
Password:
	Remember Me
	English • Español • Deutsch • Русский…

About

Help

Get Involved

Legal

Store

LJ Labs