Seeing the light: A chip in the center of this circuit board contains four lasers that convert electrical signals into light pulses. The pulses travel at high speeds along a fiber-optic link.
Intel
Replacing metal wiring with fiber optics could change everything from supercomputers to laptops.
The world of computing could change rapidly in coming years thanks to technology that replaces the metal wiring between components with faster, more efficient fiber-optic links.
"All communications over long distance are driven by lasers, but you've never had it inside devices," says Mario Paniccia, director of Intel's photonics lab in Santa Clara, CA. "Our new integrated optical link makes that possible."
Paniccia's team has perfected tiny silicon chips capable of encoding and decoding laser signals sent via fiber optics. Today, when data arrives at a computer via a fiber optic connection it has to be moved from a separate photonic device to an electronic circuit. This new system promises to speed things up because everything works in silicon.
Last week, Paniccia's team demonstrated the first complete photonic communications system made from components fully integrated into silicon chips. Electronic data piped into one chip is converted into laser light that travels down an optical fiber and is transferred back into electrical signals a few fractions of a second later. The system can carry data at a rate of 50 gigabits per second, enough to transfer a full-length HD movie in less than a second.
The silicon photonic chips could replace the electronic connections between a computer's key components, such as its processors and memory. Copper wiring used today can carry data signals at little more than 10 gigabits per second. That means critical components like the central processing unit and the memory in a server cannot be too far apart, which restricts how computers can be built.
The new Intel setup has four lasers built into its transmitter chip that shine data into a single optical fiber at slightly different wavelengths, or "colors." Chips with even more lasers should make it possible to communicate at 1,000 gigabits per second.
"Having a chip the size of your fingernail that can deliver a terabit per second changes the way you can think about design," says Paniccia. Such chips could make a big difference inside the sprawling data centers operated at great expense by Web giants like Google, Microsoft, and Facebook. "Data centers today are big piles of copper--that imposes the limits on how you arrange components inside a server," Paniccia says.
"If I could just move the memory a foot away [from the processors], I could add a whole board of memory for a single CPU instead," says Paniccia, whose team is experimenting with prototype servers to work out how to build them with photonics links inside.
Moving a server's memory away from the CPUs would also make ventilating them easier. Since roughly half the cost of running a data center, used for everything from services like Facebook to banking records, comes from cooling, that could have a significant impact.
Researchers learn how to make lasers directly on microchips—the result could be computers that download large files much more quickly.
A new kind of optical cable will provide ultrafast connections between electronic devices.
Cheap, high-performing detectors could boost Internet bandwidth and speed up computers.
Not to say that using graphene-based rather than silicon-based transistors wouldn't be a great step forward, but it strikes me that moving photons must almost always be far more efficient than moving electrons. I hope research will proceed along both these paths, but I'd bet on the fibre-optical chips described in the article....
Henri
It doesn't really matter how fast you can send bits down fiber or a wire. It's the speed of light that causes designers fits as they try and remove latency from main memory and cache accesses.
I salute Intel for their reduction in the cost structure of fiber interconnect, but the speed of light is still about a foot per nanosecond, and that won't change anytime soon. That's too slow to spread things out much and a real pain in the behind for large server designers.
Indeed. This is great news for the short term but as the industry transitions from sequential computing to massive parallelism, this is just a band-aid remedy for a huge problem that is about to reach crisis proportions in the next few years. The memory bandwidth crisis is what I call the fourth crisis and light speed communication is not even close to being the solution.
Even modest communication latency between processor and memory would not be so bad if it could be kept constant regardless of processor load and the number of parallel cores. It seems that what is needed is a new type of computer that does away with the central processor altogether. That is to say, every instruction in a program should be its own tiny parallel processor and should be directly tied to its data operands. This is not unlike the way the brain uses billions of neurons to process signals in parallel. Of course, this type of computer will require a new software model. The old algorithmic or Turing machine approach to computing will have to go.
Interesting approach indeed! I think a first step would be to enable the memory itselves to execute small sequences of distribued, stored instruction sets on it's own data. The calculations could be triggered on write, or as a response to a parameterized read command. Returning the result of a calculation, instead of the source dataset, would also reduce the bandwidth requirement of the system.
As a programmer I look forward to the challenge!
That old architecture is so widespread that it will take decades of time and dedication to produce any noticeable change. Rebuilding the processor architecture would require almost complete renovation of every programming language, which will impact every single software exist in the world. Nevertheless, replacing coppers with fiber optics will still make a significant impact, supposed the technology is indeed 'perfected'.
The concept is not new. Take the multilayer perceptron neural net that can quite easily do as you wish (in-the-end). However the greatest problem of neural nets has never been solved, although it was proposed!
Given a multilayer perceptron network you could set the values inside given you know the outputs and inputs. So therefore no need to learn the states.
No one has been able to solve this one! If you solve it, let people know please it will change the world forever.
I proposed a results orientated language/ environment a while ago to microsoft for instance, who have iq's of toe-nail clippings and couldnt understand the benefits of a result orientated programming environment.
It incorporated neural nets and learning algorithsms that could have revolutionised programming for the whole world as it produced true parallel programs with error-free, correct code. It also produced sequential programs that conformed to todays 2-d world in whatever language you wished it to be generated in. Allowing programmers to compare the computers verson against there own.
Its is a high experimental concept that required investment and feasbility studies, however Microsoft couldnt give a crap. Its run by brainless automatons.
All i can say is if you care about programming, dont talk to Microsoft, youre wasting your time.
It is call processor in memory.
There were several hurdle though.
First one was cpu and dram required opposite dielectric before. But due to miniaturization, cpu needed begin to match...ie intel atom processor.
Second is various patent some which are expiring.
Third is intel is spoiled on making huge profit on cpu. For same size, I suspect cpu is worth 10 times as much as ram.
Forth is security. Lets say there's 10 cpu/ram combo chip then there's security issue with each connection.
But there's also trend on computer on chip... so more chips with almost everything in it will be available.
Important breakthrough in Silicon Photonics
This is a very interesting breakthrough in the field of Silicon photonics. The idea of bringing "optics closer to the processor" seems more of a reality than a dream.
Intel has demonstrated each link operating at 12.5Gb/s . I would like to know what is the energy consumption per bit at such data rates? The holy grail of 100 fJ/bit or below is an important parameter for optical interconnections.
Also intel has demonstrated an optical modulator at 40 Gb/s in 2007 whose length is more than a 1mm. Is is possible to go beyond this limit of 40 Gb/s (for each link) because then we can WDM to even higher speeds , may be Tb/s ?
Um, not quite. A 50 GB BlueRay disc would have about 400-500 billion bits - transfer time about 10 seconds or more, with error correction and framing overhead. But still a BFD in consumer H/W.
Manufacturing in the United States is in trouble. That's bad news not just for the country's economy but for the future of innovation.
National Instruments has gathered customer information and data regarding some of the cost differences between building a custom solution versus using NI off-the-shelf tools. Using this data, we built the Graphical System Design ‘Build vs. Buy’ Calculator. The calculator can help show the financial differences between building a custom solution versus buying an off-the-shelf system. This paper discusses the benefits and drawbacks of both a traditional custom design approach and off-the-shelf embedded tools.
View full PDF >
Listen to story > 
On Twitter
Become a Fan on Facebook
Subscribe to the Feed
rsanchez1
213 Comments
Competing Technologies
I've read about graphene computers promising higher speeds, and I've read about these optical computers promising higher speeds. I'm no expert on computers, so can someone explain which technology is best? Maybe TR can do an article comparing the two technologies.
Reply
rocket7777
124 Comments
Re: Competing Technologies
For practical purpose they are two different things.
Optical things are kind of like network server switching type of things.
Graphine is made of one atom layer of carbon sheet. It would be like ordered cloth. Silicon based things are more like felt/paper. Anway, researchers think theoretical limit on structure is 10 times smaller than silicon.
Memoristor is variable resistor storage. It can probably replace flash drive, small harddrive, dram, and level 2 cache with it.
Reply