Moore's law and star trek

The magic smoke.

Moderators: phlip, Moderators General, Prelates

drewder
Posts: 33
Joined: Wed Oct 19, 2011 11:47 am UTC

Moore's law and star trek

Postby drewder » Tue Dec 30, 2014 12:33 am UTC

So if we assume Moore's law is true and will continue for the next few hundred years at least, how powerful would the computers on the various incarnations of the Starship Enterprise be?

User avatar
PeteP
What the peck?
Posts: 1451
Joined: Tue Aug 23, 2011 4:51 pm UTC

Re: Moore's law and star trek

Postby PeteP » Tue Dec 30, 2014 1:03 am UTC

I kinda doubt the observation can physically hold that long (you can only make something smaller to a certain point, the law doesn't require smaller so expanding the chip upwards works too for instance but we are talking about over 200 years minimum here). But if it did I doubt we could make accurate predictions about the power. Transistor counts don't translate directly to power. If we take the two year steps then in the around 250 years until we get Kirks ship we double the transistor counts 125 times leading to 4*10^37 times as many transistors. There is no telling to how much computational power that would translate. Extrapolating from current trends wouldn't do much good.
(I would be plenty happy if we got at least another ten doubling steps. *1024 is a nice number.)

wumpus
Posts: 546
Joined: Thu Feb 21, 2008 12:16 am UTC

Reality check

Postby wumpus » Fri Jan 02, 2015 4:43 pm UTC

In the 1960s (wiki says 1965) Gordon Moore observed that the amount of transistors you can cram on a chip doubled every 18 months (changed in the 1980s to two years, probably three years now).

Basically, the "law" is based on photolithography, the process of developing transistors based on a light [or other form of radiation] pattern. Intel just shipped chips at the 14nm node, and claims that the 10nm will happen (and has been quiet about previous 7nm schedules). I think flash has existed at 14nm longer (generally if you can produce memory without errors you aren't trying hard enough, fixing memory errors is well known. Fixing manufacturing errors in logic tends to involve entire spare cores.), but this seems to be transitioning to 3d forms (which can't be simultaneously crafted the same way 2d is, but seems to involve easier steps).

PeteP wrote:(I would be plenty happy if we got at least another ten doubling steps. *1024 is a nice number.)


Oddly enough, it looks like 10 doubling steps would put us right on a single silicon atom. Don't count on it. My guess is that near future chip tech will have more and more fabs bunching up near current nodes (they tend to produce more out of date stuff, but presumably the cost to produce at 16nm will drop) thus giving the appearances of the law marching on. There is also the possibility of simply increasing the size of the chips (and thus following the law as explicitly stated). AMD is supposedly producing a 500mm chip, this type of thing may be more common in the future (note that the previously mentioned issue of spare cores shouldn't be an issue: eyeballing an intel 22nm die photo shows each core about ~14mm in size, leaving plenty of room for spare cores (I doubt that cores are going to have that many more transistors. I'd even expect that future architectures will be built around making cores physically smaller).

Note that "larger chips" gets out of hand a lot faster than "smaller transistors". Even if you can manage to stack chips you will run into problems pretty fast, and stacking chips has been something of a holy grail: the same process that produces huge DRAM chips doesn't work with logic chips. Stick a DRAM chips (and use chip processes to connect thousands of wires) to a CPU and you suddenly have a massive game-changing cache. AMD claims to be doing this by next year for GPUs, I would assume that such tech could put Optrons back in business (with the possibility of plenty of bobcat cores or other shenanigans). Most "chip stacking" today means simply piling the chips on top of each other and simply wiring pins to each of them.

So "ten doubling steps" might be possible with 7nm production, 1000mm chips (>30mm per side), and stacks of 10 chips deep (think of the cooling). Hopefully we can come up with some other computer technology that isn't silicon chips by then.

A deeper question would be "what do you do with them". So far, the answer is always build it and they will come. One bright light in this is that flash seems to have unlimited demand and is least effected by this whole issue, with memory in a similar position. High performance CPU logic is pretty much stuck with the i3/5/7 cores and barely crawling ahead (with lower power chips having a mostly clear path scouted by them, but hitting the end of that scouted path realsoonnow). Stuffing more cores on the chips has worked, but customers are starting to realize that numbers of cores means less than GHz ever did (making software parallel has been "the answer" to computing since roughly MIPS released the R3000 (which had a massive cost/performance advantage to any multichip CPU), but such software has barely gotten out of the supercomputer world). GPUs appear to be utterly insatiable, but clever VR techniques could basically use current GPU tech to produce greater resolution than the eye can see (google macular eye resolution). AAAAAA-super shaded eye candy seems to be driven more by artist cost than by GPU tech. Expect the indy/megapublisher balance to get even more out of wack. I have to wonder at what point google starts producing self-analyzing flash and "ignores" off the shelf CPU tech (for values of "ignore" that involve typical google-sized purchases).

User avatar
mosc
Doesn't care what you think.
Posts: 5401
Joined: Fri May 11, 2007 3:03 pm UTC

Re: Moore's law and star trek

Postby mosc » Mon Jan 05, 2015 6:12 pm UTC

"The complexity for minimum component costs has increased at a rate of roughly a factor of two per year. " = approximately doubling the transistor count on a fixed area (quarter inch area is used in the paper) every 24 months was approximately true until about 10 years ago and isn't approximately true anymore.

CPU's getting ~2x more powerful every ~24 months has also long fallen off.

Moore's law is ether already dead or will die depending on how you define it. Only using it broadly to define something in technology growing at a stable exponential rate does it still endure.
Title: It was given by the XKCD moderators to me because they didn't care what I thought (I made some rantings, etc). I care what YOU think, the joke is forums.xkcd doesn't care what I think.

User avatar
Zamfir
I built a novelty castle, the irony was lost on some.
Posts: 7588
Joined: Wed Aug 27, 2008 2:43 pm UTC
Location: Nederland

Re: Moore's law and star trek

Postby Zamfir » Wed Jan 07, 2015 1:50 pm UTC

I thought that Moore's law in the strict sense still shows signs of life. transistor counts ( and densities) are still steadily growing near the long term trend. This graph goes to 2010 or so, and the later years show further growth yet.

It's mostly that clock speeds no longer increase together with density (and might well have to fall in the future to allow further density) . And in hindsight, those associated clock speed increases were just as important (perhaps even more important) than the raw growth in transistor counts.

With all the difficulties of EUV, we might well see decreased growth in transistor counts as well. With new fab generations becoming so expensive that people stay longer and longer on each node. But I don't think that corner has already happened with certainty. And if EUV does become commercially viable in the next few years, the transistor counts might keep growing for quite a while longer. But speed seems doomed anyway.
Spoiler:
Image
Image

User avatar
mosc
Doesn't care what you think.
Posts: 5401
Joined: Fri May 11, 2007 3:03 pm UTC

Re: Moore's law and star trek

Postby mosc » Wed Jan 07, 2015 2:09 pm UTC

Well point is you don't need to forecast to the 24th century to see the end of moore's law. Performance per watt, performance per mhz, performance per unit area, they've all fallen off. Transistor density is no longer exponential as of recently and although chips are physically larger now with better yields for a given size that means the per area density will fall off as well.

The percentage of power that is used for functional work inside the CPU is well below 50% on a modern process. Leakage power dominates. VLSI is massively about clock and power routing and management around fairly static logic for ALU's, caches, decoders, schedulers, etc. Less is different from the previous year every year we go. The belief that the semiconductor industry moves ever forward and a predictable pace is dead.

I would also argue that the cost of each node's manufacturing is growing exponentially and we're seeing fewer and fewer fabs for each node.
Title: It was given by the XKCD moderators to me because they didn't care what I thought (I made some rantings, etc). I care what YOU think, the joke is forums.xkcd doesn't care what I think.

User avatar
PeteP
What the peck?
Posts: 1451
Joined: Tue Aug 23, 2011 4:51 pm UTC

Re: Moore's law and star trek

Postby PeteP » Wed Jan 07, 2015 2:29 pm UTC

Zamfir wrote:I thought that Moore's law in the strict sense still shows signs of life. transistor counts ( and densities) are still steadily growing near the long term trend. This graph goes to 2010 or so, and the later years show further growth yet.

It's mostly that clock speeds no longer increase together with density (and might well have to fall in the future to allow further density) . And in hindsight, those associated clock speed increases were just as important (perhaps even more important) than the raw growth in transistor counts.

Yeah Dennard scaling was awesome. We can probably still get higher ghz numbers by switching to new materials. (Though I think that if you increase it by much lightspeed delays increasingly become a problem. (Much= 10X +) )

User avatar
Zamfir
I built a novelty castle, the irony was lost on some.
Posts: 7588
Joined: Wed Aug 27, 2008 2:43 pm UTC
Location: Nederland

Re: Moore's law and star trek

Postby Zamfir » Wed Jan 07, 2015 4:15 pm UTC

mosc wrote:Well point is you don't need to forecast to the 24th century to see the end of moore's law. Performance per watt, performance per mhz, performance per unit area, they've all fallen off. Transistor density is no longer exponential as of recently and although chips are physically larger now with better yields for a given size that means the per area density will fall off as well.

Yes, sure. The question is hardly "will Moore's law reach the 24th century". It's more "will some semblance of Moore's law survive to 2020".

I do find it interesting how stubborn the strict law still is, compared to vaguer versions like "computers become better at an exponential rate". The latter has clearly not been true anymore for quite some time. But exponential growth in transistor count kept going strong for quite some time, even in the face of enormous fab costs and decreasing usefulness of the extra transistors.

It seems very possible that we're currently facing the end of strict Moore growth as well, but I think the jury is still out on that. It could still be a temporary hiccup, there have always been fluctuations around the trend. We should expect a slowdown now that 192nm litho is getting stretch so ridiculously far. If EUV works well for mass production, we might see new growth again. Still at enormous cost, I bet.

wumpus
Posts: 546
Joined: Thu Feb 21, 2008 12:16 am UTC

Re: Moore's law and star trek

Postby wumpus » Thu Jan 08, 2015 2:31 pm UTC

PeteP wrote:Yeah Dennard scaling was awesome. We can probably still get higher ghz numbers by switching to new materials. (Though I think that if you increase it by much lightspeed delays increasingly become a problem. (Much= 10X +) )


I was hearing around 2000 that wire delays were the real limit in CPU speed. I don't think that would help.

Zamfir wrote:It seems very possible that we're currently facing the end of strict Moore growth as well, but I think the jury is still out on that. It could still be a temporary hiccup, there have always been fluctuations around the trend. We should expect a slowdown now that 192nm litho is getting stretch so ridiculously far. If EUV works well for mass production, we might see new growth again. Still at enormous cost, I bet.


There's a bit more room for flash with 3d flash, although I think Samsung (and presumably Micron) makes flash with even smaller transistors than Intel. I haven't seen much of a slack of demand for more flash. Presumably at some point it will be cheaper to manufacture chips at the limits, especially seeing how most of the fab won't be obsoleted in two years. Things might look a bit different if *all* your chips were manufactured at <20nm (assuming you could fit pads on them, always an issue with things like op-amps).

User avatar
mosc
Doesn't care what you think.
Posts: 5401
Joined: Fri May 11, 2007 3:03 pm UTC

Re: Moore's law and star trek

Postby mosc » Fri Jan 09, 2015 2:40 pm UTC

The thing that ended the exponential clockrate was heat and raw computational power per watt. Each node allowed for a proportional increase in clock speed without any kinds of wire delays or anything (faster speed, less distance, same latency) but it also meant more concentration of heat. Designs had to be changed in order to spread the heat out. Also, as you want to supply clock and power and ground to more and more transistors, you have more and more elaborate wiring to get them there. The more wire you use for clock, the more power (watts and waste heat) it takes to drive the chip. The PIII was quickly re-defining the amount of power a CPU would need and demanding far better coolers. AMD and Intel differed in approach for their next chips with Intel's "netburst" (basically meaning long pipelines designed for high clockrate cores) attempting to facilitate a moore's scale of clockrate. As those chips pushed upwards of 3ghz and beyond their thermal envelopes were excessive, their performance per watt terrible, and the design was mostly dumped in favor of the previous PIII architecture with multiple cores on the same die.

Parallel cores and larger caches minimizing memory use were clearer ways of delivering power without having a faster computational engine. Using extra transistors for logic in processing instructions more intelligently added efficiency with the same computational resources. Adding cache minimized memory usage which helped performance , caches now take up more room than the functional engines of CPU's. They take far less power per transistor than the "processor" parts of a chip.

The PIV killed moore's law for the clockrate of a chip. Netburst's last CPU release was on the new 65nm node in early 2006 (as well as dual CPU die variants) but the fastest arrived much earlier in Febuary 2004, Pentium 4 "Prescott" 571 at 3.8ghz. Intel would not release a faster chip until June 2014, the Core i7-4790K clocked at 4.0ghz. The i7 is massively larger in transistor count with 4 times the cores and 4 times the cache yet with less the power use of the Prescott chip.

Point is, we hit caps in exponential progress all the time with technology. No, moore's law is no different.
Title: It was given by the XKCD moderators to me because they didn't care what I thought (I made some rantings, etc). I care what YOU think, the joke is forums.xkcd doesn't care what I think.


Return to “Hardware”

Who is online

Users browsing this forum: No registered users and 5 guests