Sun Turns to Lasers to Speed Up Computer Chips

alphadogg writes to mention that Sun is attempting to move from the typical design of multiple small chips back to a unified single-wafer design. "The company is announcing today a $44 million contract from the Pentagon to explore replacing the wires between computer chips with laser beams. The technology, part of a field of computer science known as silicon photonics, would eradicate the most daunting bottleneck facing today's supercomputer designers: moving information rapidly to solve problems that require hundreds or thousands of processors."

Great idea!

I assume these systems will be water-cooled so the miniaturized sharks have somewhere to swim.

-1 : redundant

To quote Scott McNealy:

You know, I have one simple request. And that is to have SPARCS with frickin' laser beams attached to their heads!

Re:Great idea!

How do you feel about Sea Bass?

Re:Great idea!

Depends, are they ill-tempered?

Are actuators faster than direct connections?

I wonder if the time saved transmitting information via light is offset by the transition time used to translate that back into electric signals. On a single board, the distance travelled is on the order of decimeters. On a chip, micrometers. Are the time savings *that* significant? Even between peripherals, the time saved seems negligble.

Re:Are actuators faster than direct connections?

I am not an expert in electricity by no means, but I have a fundamental understanding of it (or so I think). Energy is energy. With no resistance (don't overlook this point), light traveling via laser or via electrons flowing over a wire, the speed would be the same. Now, in reality, there IS resistance... there is always a "friction" or resistance (ohm) when energy is passing over a wire. In a vacuum, a laser will move as fast as energy can possibly travel. At least on paper.

Re:

Electrons in a superconductor (a material with zero resistance) do not travel at the speed of light.

Re:Are actuators faster than direct connections?

In fact electrons in your typical electrical wire don't move anywhere near the speed of light.

http://www.eskimo.com/~billb/miscon/speed.html [eskimo.com]

Re:

No, but it depends on whether or not the receiver is current-steered or voltage-steered. If it's voltage steered then it's the propagation of the electric field that carries the signal. In which case, it can be near the speed of light.

Also, future chip-t

Re:Are actuators faster than direct connections?

When you look at a wire, or printed trace on a PCB it is not the resistance that limits how fast you can send a signal. It is inductance and capasitance that act like a low pass filter. We don't care how fact eletrons travel in wire what we care about is how fact we can change the voltage in the wire. We send data by changing voltages not by sending electrons.

Re:Are actuators faster than direct connections?

The article doesn't make it clear whether using optical communications is intended to reduce latency or increase bandwidth.

With respect to latency: the electrical signals travel at ~30% the speed of light, whereas the optical signals travel at ~70% the speed of light (it depends on refractive index, etc.). Over the distances we're talking about (as you said, mm to dm), that's only fractions of a nanoseconds delay savings [google.com]. This is on the order of a modern computer's switching time [google.com]. All this complexity to get rid of a one or two processor cycles of latency?

I suspect instead they are looking to increase bandwidth. An optical fiber can carry very high data rates. Moreover a single physical fiber can carry multiple simultaneous channels (e.g. different wavelengths of light). So the intention may instead be to create high-bandwidth links between various processors. Using on-chip lasers can make the entire assembly smaller and faster than the equivalent for electrical wires.

Really what they want, I think, is to implement the same kind of high-speed optical switching we use for transcontinental fiber-optics into a single computer or computer cluster. If you can put all the switching and multiplexing components directly onto the silicon chips, then you can have the best of both worlds: well-established silicon microchips that interface directly into well-understood high-speed optical switching systems.

Re:

It's not so much transit time, as parallelization where the big advantage is. Many frequencies of light can share the same medium without interfering with each other. Imagine many processors and memory chips streaming data to each other simultaneously,

Re:Are actuators faster than direct connections?

I don't think it's about the time it takes to transfer a single bit but the amount of bits that can be transmitted at once with light rather than wires. If we can talk line-of-sight transmission between boards, it's easy to line up an array of about a million emitters with an array of a million detectors and send back and forth the same amount of data you would need a couple thousand wires (taking translation times into account) to do.

Sun is a very entertaining company to watch. Even when their gizmos never end up in products, they are always cool.

New warning stickers...

From TFA: Each chip would be able to communicate directly with every other chip via a beam of laser that could carry billions of bits of data a second.

Do not look at chip with remaining good eye.

Whenever anyone says 50%

"This is a high-risk program," said Ron Ho, a researcher at Sun Laboratories who is one of the leaders of the effort. "We expect a 50 percent chance of failure, but if we win we can have as much as a thousand times increase in performance."

Whenever anyone says there is a 50% chance of something happening they really mean "I have no idea. No idea at all. I'm guessing."

In probability theory, "p" has a specific meaning which is roughly stated as "the ratio of the total number of positive outcomes to the total number of possible outcomes in a population". So for the number of 50% to be right, it must be known that if this research was repeated a million times, 500,000 times there would be success and 500,000 times there would be failure. But this makes no sense because the thing being measured is not a stochastic property. It is simply an unknown thing.

What is probably vaguely intended when a number like this is given is that if you took all the things in the history of the world that "felt" like this in the beginning, half of them will have worked out and half will have not.

How on earth could any mortal human know that?

But it gets even more complicated. One cannot state a probability like this without stating how confident one is in the estimate of the number. So really a person should say the probably of success of this endeavor is between 45% and 55% and this estimate will be correct 19 times out of 20.

With that as background here is what I humbly suggest 50% really means: it means "I have no idea how to quantify the error of this estimate. It doesn't matter what the estimate is because the error band could possibly stretch between 0% and 100%. So I'll split the difference and call it 50%". But that is wrong, the statement should be "I estimate the probability of success to be between 0% and 100%".

But nobody does that because it makes them look stupid.

So whenever anyone says there is a 50% chance, or a 50/50 probability of something happening, they might as well talk in made-up Klingon words, the information content of their statement will be equivalent.

Re:Whenever anyone says 50%

In probability theory, "p" has a specific meaning which is roughly stated as "the ratio of the total number of positive outcomes to the total number of possible outcomes in a population". So for the number of 50% to be right, it must be known that if this research was repeated a million times, 500,000 times there would be success and 500,000 times there would be failure. But this makes no sense because the thing being measured is not a stochastic property. It is simply an unknown thing.

This is true, if by "probability theory" you mean "Frequentism [wikipedia.org]". Frequentism is nice, for those cases where you are dealing with nice, neat ensembles. For a lot of real world situations which require probabilistic reasoning, there are no ensembles, only unique events which require prediction. For that, we often use Bayesian Probability [wikipedia.org].

Take the assertion "I'd say there's a 10% chance that there was once life on Mars." Well, from a Frequentist point of view, that's complete bullshit. Either we will find evidence of life, or we won't - either the probability is 100% or 0%. There's only one Mars out there.

In order to deal with this limitation, Bayesian Probability Theory was born. In it probabilities reflect degrees of belief, rather than frequencies of occurance. Despite meaning something quite different, Bayesian probabilities still obey the laws of probability (they sum/integrate to one, etc), thus making them mathematically compatible (and thus leading to confusion by those that don't study probability theory carefully.) Of course there are issues with paradoxes and the fact that prior distributions must be assumed rather than empirically gathered, but that does not prevent it from being very useful for spam filtering [wikipedia.org], machine vision [visionbib.com] and adaptive software [norvig.com].

As someone who professionally uses statistics to model the future performance of a very large number of high-budget projects at a major U.S. defense contractor, I can assure you that his statement was much more in line with the Bayesian interpretation of probability than the Frequentist view you implicitly assume.

Sorry for the rant, I just get very annoyed when people assume that Frequentism is all there is to statistics - Frequentism is just the beginning.

But it gets even more complicated. One cannot state a probability like this without stating how confident one is in the estimate of the number.

Of course! But where did the confidence interval come from, and how much confidence do we have in it? It's important to provide a meta-confidence score, so that we know how much to trust it! That too, however, should be suspect - indeed even moreso because it is a more complex quantity to measure! So a meta-2 confidence score is in order, for any serious statistician... But why stop there?!

With that as background here is what I humbly suggest 50% really means: it means "I have no idea how to quantify the error of this estimate. It doesn't matter what the estimate is because the error band could possibly stretch between 0% and 100%. So I'll split the difference and call it 50%".

So, if someone does not give an error bound on an estimate, we should assume that the error is maximal?

So whenever anyone says there is a 50% chance, or a 50/50 probability of something happening, they might as well talk in made-up Klingon words, the information content of their statement will be equivalent.

Or, it's entirely possible that that 50% number is somewhat accurate, because they know something about the subject that you do not.

Not about single wafer design

If I understood correctly this is not about single wafer design but exactly the opposite: regaining the speed of 'single wafer design' with multiple chips by using optical communications between chips increasing the inter-chips bandwidth (normally intra-chip bandwith is much higher than inter-chip bandwith so this is a bottleneck).

Re:

A better link providing more information:
http://www.theinquirer.net/gb/inquirer/news/2008/03/24/sun-silicon-photonics-macro [theinquirer.net]

link to the original story (!)

Why, why, why do people submit second-hand links to Slashdot?

The byline of the Seattle Times story is "John Markoff New York Times". 5 seconds with Google's site:nytimes.com reveals the original story [nytimes.com] with better explanation and more quotes from Sun personnel.

A really high bridge

On chip they are pumping the signal over a traces with mm range lengths and um range widths, off chip it's over traces with dm range lengths and mm range widths. Timing and power consumption are hard enough problems on chip, off chip they become much harder ... not to mention that most of the power consumed either goes into EM or gets coupled into other signals.

Serial connections help with the timing, but do diddly for power and noise. That's where optical comes in.

Re:Why not...

To use the beloved transportation analogy: it's like moving your cargo off of trucks and onto a high-speed train. Yes it takes time to move cargo, but it's worth it if the time savings of the high-speed train are big enough (for long enough distances, the savings can be significant).

In this case, there may be a delay associated with signal processing, but if the optical transmission is sufficiently faster than an equivalent electrical one, then it's worth it. Considering that electrical signals themselves need to undergo various kinds of switching and processing anyway (data written or read from a bus), I don't know that converting to laser signals will add much of a delay.

Re:

So do transistors. What's your point ? Analog computation ? Yurk.

Re:

Don't worry, someone will ask it a question that is a paradox before then, and the whole thing will destroy itself with sparks and slowed audio.

Re:

I don't know if this is a serious question or not, but one assumes that the lasers will operate in completely sealed environments (e.g. inside an IC package) or over optical fibers if they need to traverse free space. I think the intra-package situation i

Re:Don't Shake the computer!

how will dust be solved?

Why don't you crack open your 3.5" hard disk drive and find out why dust doesn't bother those sensitive platters? ;)