7 ms·

How low (power) can you go?

98 points by cstross 14y ago

codedivine 14y ago
I think the assumption that Moore's law, or even Koomey's law, is going to hold for another two decades is wildly optimistic. I would expect a taper off much sooner, perhaps even this decade. The physics and the economics of smaller process nodes is becoming harder and harder, to the point where anything beyond 10nm or so looks extremely difficult.
- eleitl 14y ago
  Indeed, see http://www.semiwiki.com/forum/content/1388-scariest-graph-i-ve-seen-recently.html http://www.semiwiki.com/forum/content/1388-scariest-graph-i-...
  - Someone 14y ago
    That graph ignores that there will be pressure to go to smaller sizes not to get more power for the same size, but to get the same power for less power usage. On the other hand, I would like to see a graph plotting time or nanometers vs cost of the factory needed to make them. I cannot find one, but I remember seeing ones that hinted that the cost of making a production line would become prohibitive for everyone.
- Dn_Ab 14y ago
  I don't. The slowing of the pace of innovation has been claimed for centuries but has yet to happen. Switching architectures, materials, moving to atomic scale - there is still plenty of room at the bottom. I'm excited for memrisotor tech myself.
eleitl 14y ago
Apparently, Charlie hasn't gotten the memo: http://www.semiwiki.com/forum/content/1388-scariest-graph-i-ve-seen-recently.html http://www.semiwiki.com/forum/content/1388-scariest-graph-i-... These are financial limits, the scaling limits will end at about 1-3 nm. But it will take longer, as doubling will no longer in constant time.
- Symmetry 14y ago
  Well, that does imply that we could keep making the same processors at smaller sizes for the same price. Power density issues might constrain speeding things up at that point, but it would certainly be a continuation of Koomey's law. EDIT: Maybe more likely is that we might be seeing a future slowdown of Moore's law to a doubling every 24 months or more, allowing fabs more time to amortize their investments. This would be bad, but not the end of the world.
yaakov34 14y ago
A couple of notes about the limits of things that have already been reached today: A pixel in a camera sensor has to be a couple of wavelengths in size. This is basically the size that we can make them now. There are also diffraction limits for any imaging system, and a size limit on the lens (if your lens is too small, its surface won't collect enough photons in low light to make an image - no matter how you focus them). So I don't think it would be possible to make a deep sub-millimeter camera with usable image quality, and I don't think the 10-order-of-magnitude gains in power efficiency will be coming, either (if you can't shrink it, you're stuck switching larger circuitry and dealing with larger charges). Now, spy cameras can already be annoyingly tiny, but I don't think we're headed for cameras on a speck of dust or some such. One piece of evidence for this is that tiny sub-millimeter insects have lost most or all of their vision - eyes don't scale well to these sizes. Genome sequencing: I think there are limitations to what can be done with these that are already being hit. E.g. looking at the USB-attached genome reader that Charlie Stross mentions, I see that it really needs to have the material which you're feeding it to be separated into DNA/non-DNA fractions for decent performance (otherwise, the micropores will be deluged by non-DNA material), and it needs a conventional chemistry lab in order to cut something as large as a human genome into manageable chunks. For this you need things like centrifuges, which can't become microscopic. I think there are fundamental limits reached here as well, since you can't make a pore arbitrarily small (molecular-sized) without having all sorts of molecular junk jam into it, if it is present.
- Dn_Ab 14y ago
  He was perhaps over enthusiastic with the sequencing aspect and while I don't think it will come about as he describes (for the reasons you give), I expect some method of real time ecosystem analysis to exist. Your point on camera sensor size is well made. But compressed sensing will allow us to get much lower energy usage and exploited redundancies at local nodes in the network will allow us to do well at otherwise impractical sizes.
  - yaakov34 14y ago
    I am not sure what you mean by compressed sensing - if you're only interested information such as "child-sized object moving rapidly across field of view", then you can get away with smaller sensors, but here, CS is also talking about faces being recognized in a crowd. That's going to require a relatively bulky camera system, just because you need to collect a large amount of photons (which requires an aperture/lens of certain size), and record them on a sensor which can decide which direction they came from (also needs a certain size). This is the reason why flies and bees have eyes that are (small) webcam-sized, and they probably can't pick out facial features, even if they cared to or had the brain to process them. For the sequencing, I am really a lot less sure what's ultimately possible, but I wanted to call attention to possible overhyping. If you read the manufacturer's advertisement for the USB genome scanner, it starts out by saying that it accepts a drop of saliva or blood and starts producing a genome. Then it sort of segues into saying that you'd want to separate out the chromosome which you want sequenced (which involves centrifuging the cells to separate the chromosomes by size), and cut that chromosome into manageable pieces (wet chemistry, more centrifuging), and then it will read most of that chromosome in its 6-hour lifetime, but of course you'll only have piecewise sequences to put together. Uh, that was not what you said in the beginning, and I have a feeling there are more gotchas - this is, after all, still just the advertisement.
lsiebert 14y ago
Either processors could share cameras, or each processor could have a low resolution camera that works in aggregate, perhaps similar to how a fly's eyes work. Or we could have both. Computer vision doesn't have to replicate human vision to be useful, especially if the processing is fast enough.
ctdonath 14y ago
I've occasionally mulled what a 1-bit camera could do, in the same notion as a 1-bit audio A/DC. Cross this with CS's gazillion millimeter computers and emerging 3D image extraction (term escapes me, take lots of photos and deduce the source objects a la Google 3D Maps). One camera may be too small for an image (both resolution & color depth). A lot of them, thrown randomly around, may provide interesting results given enough math.
- eru 14y ago
  See http://dsp.rice.edu/cscamera http://dsp.rice.edu/cscamera for an almost practical single pixel camera.
eru 14y ago
Nobody mentioned that Moore's first law isn't just about the number of transistors doubling. It's about the number in the chip that's cheaper per transistor. Also there's Moore's second law about the cost of fabs rising exponentially at about the same rate. If GDP growth doesn't accelerate dramatically, we won't be able to afford fabs any longer in a few decades.
wwalker3 14y ago
Stross' comparison between Cray 1 performance and that of a modern smartphone seems off. He says: "A regular ARM-powered smartphone, such as an iPhone 4S, is some 12-13 orders of magnitude more powerful as a computing device than a late 1970s-vintage Cray 1 supercomputer." A Cray 1 could peak at about 250 MFLOPs/s (http://en.wikipedia.org/wiki/Cray-1 http://en.wikipedia.org/wiki/Cray-1), and a modern smart phone like the Galaxy Nexus peaks at about 9.6 GFLOPs/s (using ARM Neon instructions on both cores). That's less than two orders of magnitude difference. Floating-point power efficiency seems to have improved by about 6-7 orders of magnitude in that time though, which is very nice :)
- lmm 14y ago
  Maybe he's comparing integer operations? ARM has relatively underpowered floating-point because most of the uses for an ARM (at least historically) don't involve it.
  - maxerickson 14y ago
    He notes it is a mistake in the comments. But the details of the comparison don't change the part where a modern cell phone is nicely comparable to a $millions computer from a few decades ago.
pressurefree 14y ago
fluorescent LEDs

98 points by cstross 14y ago

I think the assumption that Moore's law, or even Koomey's law, is going to hold for another two decades is wildly optimistic. I would expect a taper off much sooner, perhaps even this decade. The physics and the economics of smaller process nodes is becoming harder and harder, to the point where anything beyond 10nm or so looks extremely difficult.

eleitl 14y ago
Indeed, see http://www.semiwiki.com/forum/content/1388-scariest-graph-i-ve-seen-recently.html http://www.semiwiki.com/forum/content/1388-scariest-graph-i-...
- Someone 14y ago
  That graph ignores that there will be pressure to go to smaller sizes not to get more power for the same size, but to get the same power for less power usage. On the other hand, I would like to see a graph plotting time or nanometers vs cost of the factory needed to make them. I cannot find one, but I remember seeing ones that hinted that the cost of making a production line would become prohibitive for everyone.
Dn_Ab 14y ago
I don't. The slowing of the pace of innovation has been claimed for centuries but has yet to happen. Switching architectures, materials, moving to atomic scale - there is still plenty of room at the bottom. I'm excited for memrisotor tech myself.

eleitl 14y ago

Apparently, Charlie hasn't gotten the memo: http://www.semiwiki.com/forum/content/1388-scariest-graph-i-ve-seen-recently.html http://www.semiwiki.com/forum/content/1388-scariest-graph-i-... These are financial limits, the scaling limits will end at about 1-3 nm. But it will take longer, as doubling will no longer in constant time.

Symmetry 14y ago
Well, that does imply that we could keep making the same processors at smaller sizes for the same price. Power density issues might constrain speeding things up at that point, but it would certainly be a continuation of Koomey's law. EDIT: Maybe more likely is that we might be seeing a future slowdown of Moore's law to a doubling every 24 months or more, allowing fabs more time to amortize their investments. This would be bad, but not the end of the world.

yaakov34 14y ago

A couple of notes about the limits of things that have already been reached today: A pixel in a camera sensor has to be a couple of wavelengths in size. This is basically the size that we can make them now. There are also diffraction limits for any imaging system, and a size limit on the lens (if your lens is too small, its surface won't collect enough photons in low light to make an image - no matter how you focus them). So I don't think it would be possible to make a deep sub-millimeter camera with usable image quality, and I don't think the 10-order-of-magnitude gains in power efficiency will be coming, either (if you can't shrink it, you're stuck switching larger circuitry and dealing with larger charges). Now, spy cameras can already be annoyingly tiny, but I don't think we're headed for cameras on a speck of dust or some such. One piece of evidence for this is that tiny sub-millimeter insects have lost most or all of their vision - eyes don't scale well to these sizes. Genome sequencing: I think there are limitations to what can be done with these that are already being hit. E.g. looking at the USB-attached genome reader that Charlie Stross mentions, I see that it really needs to have the material which you're feeding it to be separated into DNA/non-DNA fractions for decent performance (otherwise, the micropores will be deluged by non-DNA material), and it needs a conventional chemistry lab in order to cut something as large as a human genome into manageable chunks. For this you need things like centrifuges, which can't become microscopic. I think there are fundamental limits reached here as well, since you can't make a pore arbitrarily small (molecular-sized) without having all sorts of molecular junk jam into it, if it is present.

Dn_Ab 14y ago
He was perhaps over enthusiastic with the sequencing aspect and while I don't think it will come about as he describes (for the reasons you give), I expect some method of real time ecosystem analysis to exist. Your point on camera sensor size is well made. But compressed sensing will allow us to get much lower energy usage and exploited redundancies at local nodes in the network will allow us to do well at otherwise impractical sizes.
- yaakov34 14y ago
  I am not sure what you mean by compressed sensing - if you're only interested information such as "child-sized object moving rapidly across field of view", then you can get away with smaller sensors, but here, CS is also talking about faces being recognized in a crowd. That's going to require a relatively bulky camera system, just because you need to collect a large amount of photons (which requires an aperture/lens of certain size), and record them on a sensor which can decide which direction they came from (also needs a certain size). This is the reason why flies and bees have eyes that are (small) webcam-sized, and they probably can't pick out facial features, even if they cared to or had the brain to process them. For the sequencing, I am really a lot less sure what's ultimately possible, but I wanted to call attention to possible overhyping. If you read the manufacturer's advertisement for the USB genome scanner, it starts out by saying that it accepts a drop of saliva or blood and starts producing a genome. Then it sort of segues into saying that you'd want to separate out the chromosome which you want sequenced (which involves centrifuging the cells to separate the chromosomes by size), and cut that chromosome into manageable pieces (wet chemistry, more centrifuging), and then it will read most of that chromosome in its 6-hour lifetime, but of course you'll only have piecewise sequences to put together. Uh, that was not what you said in the beginning, and I have a feeling there are more gotchas - this is, after all, still just the advertisement.

lsiebert 14y ago

Either processors could share cameras, or each processor could have a low resolution camera that works in aggregate, perhaps similar to how a fly's eyes work. Or we could have both. Computer vision doesn't have to replicate human vision to be useful, especially if the processing is fast enough.

ctdonath 14y ago

I've occasionally mulled what a 1-bit camera could do, in the same notion as a 1-bit audio A/DC. Cross this with CS's gazillion millimeter computers and emerging 3D image extraction (term escapes me, take lots of photos and deduce the source objects a la Google 3D Maps). One camera may be too small for an image (both resolution & color depth). A lot of them, thrown randomly around, may provide interesting results given enough math.

eru 14y ago
See http://dsp.rice.edu/cscamera http://dsp.rice.edu/cscamera for an almost practical single pixel camera.

eru 14y ago

Nobody mentioned that Moore's first law isn't just about the number of transistors doubling. It's about the number in the chip that's cheaper per transistor. Also there's Moore's second law about the cost of fabs rising exponentially at about the same rate. If GDP growth doesn't accelerate dramatically, we won't be able to afford fabs any longer in a few decades.

wwalker3 14y ago

Stross' comparison between Cray 1 performance and that of a modern smartphone seems off. He says: "A regular ARM-powered smartphone, such as an iPhone 4S, is some 12-13 orders of magnitude more powerful as a computing device than a late 1970s-vintage Cray 1 supercomputer." A Cray 1 could peak at about 250 MFLOPs/s (http://en.wikipedia.org/wiki/Cray-1 http://en.wikipedia.org/wiki/Cray-1), and a modern smart phone like the Galaxy Nexus peaks at about 9.6 GFLOPs/s (using ARM Neon instructions on both cores). That's less than two orders of magnitude difference. Floating-point power efficiency seems to have improved by about 6-7 orders of magnitude in that time though, which is very nice :)

lmm 14y ago
Maybe he's comparing integer operations? ARM has relatively underpowered floating-point because most of the uses for an ARM (at least historically) don't involve it.
- maxerickson 14y ago
  He notes it is a mistake in the comments. But the details of the comparison don't change the part where a modern cell phone is nicely comparable to a $millions computer from a few decades ago.

pressurefree 14y ago

fluorescent LEDs