Increasingly lower energy computing is obviously happening, but there are lower bounds to how little power you need for a useful device, it would seem. Wireless networking will always require a reasonable amount of power to achieve any reasonable distance/reliability.
The rest of the comments are as well, especially everyone's predictions for what they'll see in 10 years in computing. Quite interesting looking back in time of roughly a computer century and seeing the viewpoints we had back then.
The point is, within ten years, we won't be using silicon-based computers. They'll be made obsolete by DNA/protein type bio-computers or maybe molecular computers.
- by MonkeyMan
Do we geeks always sound this sure of ourselves when we're dead wrong?
Yes we do, but this feels extra-wrong/possibly-a-joke because I don't remember any actual research into-- or any reason besides the future to believe in-- "bio-computers" or "molecular computers" in '00.
Wow, those comments make people from the year 2000 seem like terrible people...
I like how the article references moores law, which I guess is supposed to be the doubling of transitors... this is predicting a doubling in clock speed.
Even better yet is the Starfire movie made by Sun Microsystems in 1993. I'm still waiting for my desk that is a display and scanner.
http://www.youtube.com/watch?v=NKJNxgZyVo0
Edit: I found my link the source page, with a longer downloadable version and "director's cut" of the film: http://www.asktog.com/starfire/
Do we not have 10ghz processors? My impression is that they exist, or you can at least overclock to that level, it's just neither reliable nor economical, and real workloads that require that much horsepower can by and large be parallelized, so the difference between quad-core 3.5 ghz and 10ghz is mostly academic.
We have 10GHz digital logic for sure (Thunderbolt drives its differential lines that fast, I believe). You can do that for a small region of a chip with a handful of transistors. But you can't run a whole 50+ mm^2 die at that speed for the simple reason that it can't be cooled by non-exotic methods. You just can't get the heat out fast enough, so you can't ship the product.
So for now, with ambient temperature cooling, the best modern chips top out at about 5GHz. That will likely improve slowly with each process generation due to efficiency improvements, but it's not going to change much.
If you can run a quad-core chip at 3 GHz, then there it is almost certainly feasible to run one core at 10 GHz.
I suspect the real limit is in the size of the L1 and predecode caches. I expect they are limited by propagation delay, not transistor speed. The only way to get more cache is to use more cores.
Heat transfer doesn't scale quite like that. Once the surface area gets beyond "tiny subcircuit" you're limited by the essentially 1D transfer "up" through the packaging and can't rely on the ability of the cooler silicon in the surrounding regions. One core is still a really big circuit.
And yes, caches are big regions and hard to make faster synchronously. But you can treat this with pipelining -- that's one reason even the 32k L1 data cache in SNB/IVB has a 4 cycle (!) latency. Note that the even the gargantuan L3 cache on these chips (which is literally 1/3 of the die area) is still running at full speed, just with a ~30 cycle latency.
>Looks great, but ignores the fact that transistors don’t scale like they used to. Remember, the point of near-threshold voltage and the research into replacing silicon is intended to move the bar forward bit by bit, not to re-enable the classic Dennard scaling of the 1980s and 1990s. That era is gone, and nothing short of a miracle material that fulfills all the roles of silicon will ever bring it back.
Incremental changes in architecture do not have to equate to incremental changes in capability. Replacing silicon with a different substrate can introduce new time complexities for old problems. For example, mapping neural models onto memristors could scale better than mapping same models onto traditional silicon. Mapping quantum physics models onto qbits will scale better than mapping same models onto silicon. Mapping protein folding onto protein based computers could... and so on.
I've wondered if mobile devices will reach the point of such low power consumption that they can be powered by all the radio wave energy already enveloping us.
I think a far more likely senerio, and probably more likely is that our everyday movements will charge much like the shake flashlights you see from time to time.
I'm afraid that won't be enough. There's a lower limit on how much you can compute with a given amount of energy. If I recall correctly, switching a bit will cost at least ln(2)kT energy, or about 310^-21 joules at room temperature. According to Wikipedia[1], FM signals are minimally 10^-18 watt. If we have about a hundred antennas and assume a somewhat larger signal strength, we could thus viably use 10^-15 watt. Sadly, that will only allow us to switch 400000 bits per second[2], not nearly enough for any meaningful computation, let alone communication. Notice that I'm not talking about storing a few hundred thousand bits, I'm talking about switching them from 0 to 1 or vice versa. Adding two numbers will probably switch dozens of bits. And that is with equipment operating at the limit of what is possible by the laws of physics. So likely we'll end up a few factors thousand lower. A few hundreds of bits per second is probably enough for some RFID like system, but not for anything useful.
Accurately measuring geospatial location via GPS, making a phone call, or playing a game is meaningful.
Starting with a definition like that, it's not hard to see how the author concludes that near-zero-energy computing isn't possible. But a more modest definition-- for example, of compute-enabled "smart" versions of already-existing products-- may make that vision possible. Thinking of sensors as a form of meaningful computation expands the range even more.
For example, I've seen switches that harvest enough energy from the act of pressing them, to communicate their change in status to the controlled device. Now consider coupling that energy to some logic. For example, maybe a single light switch dynamically determines which of several lights you want to switch on or off in the space.
The real barrier to "ubiquitous meaningful low-energy computation" is that the marginal extra energy for adding computation to an existing device must be small compared to the energy that device already draws. The classic example of this is automobiles, which have been acquiring more and more sensors and internal control logic over the years. As logic components get to lower energy consumption, why shouldn't those possibilities jump to other devices?
What happened to the magical PixelQi: http://pixelqi.com/ screens that we have been promised for the last 3 years? Supposedly completely reflective full color fast refreshing LCD screens with very low power consumption. Seems that all that's available is a screen that you have to hack into a very limited number of netbook models yourself. Low power screens would make a far greater impact than any advances in low power processing. Can't understand why Apple or somebody else isn't all over these.
Even if the color reproduction is terrible, imagine the benefits of being able to write code in the park. We'd be seeing a lot of very tan developers.
screens are too power hungry. The evolution is toward glasses that project light on you eye. The first steps were to remove keys and wires, the next one is to remove screen and speakers.
I think it's more the backlight is power hungry, not the display itself. Direct projected light into the eyes seems to have the risk of probably causing vision problems.
I remember reading a lecture by Feynman where he discusses reversible computing and suggests that you can get processing down to insanely low power usage using those methods.
I think the author may have misunderstood the claim. Intel was probably referring literally to low power computing (what happens inside Intel's chips) not to low power any-thing-you-can-do-with-a-computer(display, communicate with others, etc). In other words the processor.
Nearly all the power and heat problems in processors has to do with impedance mismatches between materials in the circuit. It's been about 5 years since I went to a conference called Beyond Moore's Law, but I remember a brilliant talk on a 5-6 order of magnitude decrease in power that is possible though impedance matching. (I couldn't find a link online, sorry!)
I suspect (rather arrogantly, since I have not seen intel's article directly) that this is what Intel was talking about.
I remember seeing videos of some lectures by Hal Abelson on a project he was working on (in scheme of course) about a network of processing units sharing informations and distributing computations.
33 comments
[ 3.4 ms ] story [ 75.8 ms ] threadhttp://www.geek.com/articles/chips/intel-predicts-10ghz-chip...
Increasingly lower energy computing is obviously happening, but there are lower bounds to how little power you need for a useful device, it would seem. Wireless networking will always require a reasonable amount of power to achieve any reasonable distance/reliability.
Do we geeks always sound this sure of ourselves when we're dead wrong?
I like how the article references moores law, which I guess is supposed to be the doubling of transitors... this is predicting a doubling in clock speed.
Edit: I found my link the source page, with a longer downloadable version and "director's cut" of the film: http://www.asktog.com/starfire/
So for now, with ambient temperature cooling, the best modern chips top out at about 5GHz. That will likely improve slowly with each process generation due to efficiency improvements, but it's not going to change much.
I suspect the real limit is in the size of the L1 and predecode caches. I expect they are limited by propagation delay, not transistor speed. The only way to get more cache is to use more cores.
And yes, caches are big regions and hard to make faster synchronously. But you can treat this with pipelining -- that's one reason even the 32k L1 data cache in SNB/IVB has a 4 cycle (!) latency. Note that the even the gargantuan L3 cache on these chips (which is literally 1/3 of the die area) is still running at full speed, just with a ~30 cycle latency.
Wow. I did not know the L3 cache ran so fast.
Doesn't really seem practical to run a datacenter using the methods in the video however. But, it's still possible.
Incremental changes in architecture do not have to equate to incremental changes in capability. Replacing silicon with a different substrate can introduce new time complexities for old problems. For example, mapping neural models onto memristors could scale better than mapping same models onto traditional silicon. Mapping quantum physics models onto qbits will scale better than mapping same models onto silicon. Mapping protein folding onto protein based computers could... and so on.
1: http://en.wikipedia.org/wiki/Orders_of_magnitude_(power)#att... 2: http://www.wolframalpha.com/input/?i=1+bit+%2F+%28ln%282%29+...
http://en.wikipedia.org/wiki/Crystal_radio
Starting with a definition like that, it's not hard to see how the author concludes that near-zero-energy computing isn't possible. But a more modest definition-- for example, of compute-enabled "smart" versions of already-existing products-- may make that vision possible. Thinking of sensors as a form of meaningful computation expands the range even more.
For example, I've seen switches that harvest enough energy from the act of pressing them, to communicate their change in status to the controlled device. Now consider coupling that energy to some logic. For example, maybe a single light switch dynamically determines which of several lights you want to switch on or off in the space.
The real barrier to "ubiquitous meaningful low-energy computation" is that the marginal extra energy for adding computation to an existing device must be small compared to the energy that device already draws. The classic example of this is automobiles, which have been acquiring more and more sensors and internal control logic over the years. As logic components get to lower energy consumption, why shouldn't those possibilities jump to other devices?
The following, related, analysis was posted to HN a while ago: http://www.antipope.org/charlie/blog-static/2012/08/how-low-...
Even if the color reproduction is terrible, imagine the benefits of being able to write code in the park. We'd be seeing a lot of very tan developers.
Also "a major partnership with 3M": http://www.pixelqi.com/press/3m_new_ventures
Me, I'm waiting for color e-paper. But I'm not holding my breath.
The fast refreshing fully reflective full color tech is called mirasol.
I couldn't find a link to the lecture in question, but here is some recent research to give an overview - http://ercim-news.ercim.eu/en79/special/micropower-towards-l...
Nearly all the power and heat problems in processors has to do with impedance mismatches between materials in the circuit. It's been about 5 years since I went to a conference called Beyond Moore's Law, but I remember a brilliant talk on a 5-6 order of magnitude decrease in power that is possible though impedance matching. (I couldn't find a link online, sorry!)
I suspect (rather arrogantly, since I have not seen intel's article directly) that this is what Intel was talking about.