Slashdot Mirror
Supercomputer To Use Optical Router
Izmunuti writes "From a NYTimes article: 'Highlighting a radical departure in the design of the fastest computers, the California Institute for Telecommunications and Information Technology plans to announce on Monday that it will use an optical router
designed by a Texas company as the heart of a
campus-wide supercomputer that will be woven together with optical fibers.'"
At first we wanted everything to be transfered from analog type formats to pure digital, now we are exploiting analog formats (like optics) to get insane speeds.
Will we go back to all digital in about 10 years?
When it comes down to it, the computers do the work. You can do useful supercomputing with almost no networking, you can't do useful supercomputing with blindingly fast networks and no computers.
(Somehow, the quote reminds me of people who think that managers and lawyers are the important part of a company, and engineers and customer service are a nuisance to be minimized.)
Anyway, we're about pushing the limits of copper, with 1000bT, and I'd imagine network speeds will only continue to climb with increased use of fiber. I can see, in 5 to 10 years, optical switches becoming more common in office environments as file sizes and network speeds continue increasing.
Just as the article started to pique my interests, it was over. That sucks!
Yet another technology article written without any real information. I realize in writing you are supposed to write to the common reader, but sometimes it seems like they would be better off not writing about it at all if they didn't intend on clueing us in on any of the facts.
A slip of the foot you may soon recover, but a slip of the tongue you may never get over. -Benjamin Franklin
This would be a network of computers, not a supercomputer. The definition is becoming more lenient, so in a few years everyone on the internet will be a node in the world's largest (and only) supercomputer, and 80% of them will be redundantly running through Windows DLLs. Yay.
I don't know how he calculates that. Maybe that matches the raw number of logic operations of a human brain, but a digital computer has a completely different organization, so it's like comparing apples and elephants.
The brain's advantage comes through the fact that the "logic" is embedded within and mixed up with an incredibly powerful fully associative storage system. The keys and values aren't little byte strings or numbers like digital computers use, but instead they are high-level concepts and experiences. We don't even know how to begin designing a direct emulation of this kind of hardware.
OTOH, it might take someone 10 minutes to manually do a long division problem that the computer can solve in under 1 nanosecond. However, even with all of the awesome math throughput provided by supercomputers that consume tens of kilowatts of power, nobody's come up with a system that has the real-world common sense and precise realtime control capablities of a 1 milliwatt cockroach brain. (Did you know that they can fly? I discovered that one day by spraying one on the ceiling. Scared the living shit out of me.)
Obviously, making speed comparisons between brains and digital computers is utterly meaningless when the fundamental operations they perform are so completely different.
There is still some room to grow left in the semiconductor business....but not much. Now that chip makers are near the limit in the "how small can I shrink it race", the goal now is to fit more die on single silicon wafer. As the "need for speed becomes more critical I think optical computing breakthroughs will become more common place as soon as the semiconductor business and research community start to move away from silicon semiconductors. When the focus on optic technology becomes dominant silicon chips won't go away, but they won't be as high tech.
"You helped our nation celebrate its bicentennial in 17 -- 1976." --George W. Bush, to Queen Elizabeth, Wash
Actually, what we need is a large broadcast multimode fiber network with tunable transmitters and receivers, no switching involved.
As a paper I once read.. I think it was called "The Fibersphere".
Switching is simply a hack to get around a lack of bandwidth.
The concept is that if we have this large, broadcast fiber network, and tunable receivers sensitive enough, everyone could transmit and receive on the same thing, and talk to anyone else. Tuning in to the right signal is all that would be required... just like RF in the atmosphere.. but with much, much higher bandwidth.
Problem? WE don't have tunable laser emitters yet.
I agree with what you are saying, but what you describe is an optical switch not an optical router.
;-)
(a switch switches circuits or light channels in this case and a router routes packets).
I read trough their website (www.chiaro.com) but wasn't clear on how they can identify the destination addresses of the packets (essential for routing) without some sort of photonic-electrical conversion. Then it won't be an all optical router, would it?
cheers,
Frank
A truth that's told with bad intent, Beats all the lies you can invent. -- William Blake
Actually, there are lots of tunable lasers, and the diode lasers used in communications are intrisically tunable because the refractive index of the semiconductors used to make diode lasers depends on temperature. Thus to tune the emission wavelength over the laser diode's range, you only have to make it run a little hotter or a little cooler.
I think the bigger problem it that it's difficult and expensive to build electronics to demodulate extremely high frequency signals, so you'd only rather have a few of those expensive boxes feed a bunch of cheap boxes instead of having to give everyone the expensive box.
"It take 9 months to bear a child, no matter how many women you assign to the job."
Well... thats half true. Certainly more bandwidth, but not normally lower latency.
Light through vacuum is quick, light through glass isn't as quick. Couple this with the inability for the light to travel in a straight line through the fibre (it bounces around off the sides -- more or less). Electrical signals through copper don't experience these affects as much.
Lastly, a rather complex and heavily delayed circuit has to convert the electrical signals to light, and back again. This takes time -- but the percentage of time taken is small in comparison to normal travelling distances but don't expect them to make a slowed down PCI bus using fibre any time soon.
Rod Taylor
That's the lamest excuse i've ever heard. If you don't like their policies, read the news somewhere else instead of polluting their database. They don't send you spam anyway.
Unfortunately light travels pretty slowly.
.. regardless of bus speed ..guess what ... it's a bottleneck!
A 1 Ghz processor does a computation before light can travel 35 centimeters (little more than one foot). So if memory or a second cpu is placed more than a foot away