Comments Locked

42 Comments

Back to Article

  • dasHHa - Friday, July 30, 2010 - link

    Moore's law still going on!
  • Muscles - Friday, July 30, 2010 - link

    A lot to look forward to in the future.
  • DJMiggy - Friday, July 30, 2010 - link

    Very good article. VERY exciting stuff. I look forward to the future. Both near and far!

    *sings* In the year two thousaaaaaaaaaaaaaaaaaaand!
  • GullLars - Wednesday, August 4, 2010 - link

    *sings* In the year 2525...

    I just drifted off and almost drooled thinking about this combined with tiered non-uniform Solid State Storage (L1-3/4 volatile, and 4/5 none-volatile, highly parallel).

    Already today you can get 100+ MB/s (read) from a single NAND chip (ONFI 2.x), so scaling out to 2^n channels you could quickly aggregate bandwidth to several GB/s for parallel workloads or bulk (>1MB) transfers, without insane costs.

    With a re-make of storage management (like Fusion-IO's new "Virtual Storage Layer") and streamlining of the initiating part of program codes, you could load the operating system and any program near-instant. I've already had loading times around a second for many programs (and a few seconds for the really heavy ones) for two years now thanks to my RAIDed SSDs <3, but this would be a new level of performance.
  • KG Bird - Friday, July 30, 2010 - link

    This was a very interesting piece and reminds me of a few things I'd stuffed into the back of my brain. Just to clear up a couple of things though...

    From Wikipedia: Epitaxy refers to the method of deposition a monocrystalline film on a monocrystalline substrate.

    So this device is made through an epitaxial process.

    Intel puts down an epitaxial buffer layer to smooth out the differences in lattices between the silicon and the next layer. If you don't use a buffer layer, you get defects such as voids or bulges that lead to the undesirable properties mentioned, or it just plain overheats at the defect location and doesn't work.
  • 3DoubleD - Friday, July 30, 2010 - link

    It is likely that the buffer layer is a metamorphic layer. The use of a quaternary alloy (AlInGaAs) alloys engineers to slowly vary the lattice constant from that of Si to that of InP. This prevents the lattice strain from ever reaching a point where it is energetically favorable to create a mismatch dislocation and thus degrading device performance. All is easier said that done and it's great that Intel was able to successfully implement this technique.

    Another method of integrating high performance III-V semiconductor materials onto Si are nanowires. The small nanowire diameters allows for mismatched materials to be connected without metamorphic layers.

    In the next 10 years we will finally begin to leave the silicon dominated electronics industry into this hybrid type, where the superior performance of III-V semiconductors will meet the affordability and level of integration only enjoyed by the silicon community. There is no telling what this will look like, but integrated photonic circuits is just the beginning. Everything from a high mobility electron transistor to a single electron transistor would become possible on silicon in an affordable way. This will extend outside the integrated circuit world and into other areas such as solar cells, photo detectors, LEDs, biosensors, ect. We are only scratching the surface.
  • eanazag - Friday, July 30, 2010 - link

    Good article. I am waiting for some free light Intel thingys. It happens to be 2010 now.
  • zalves - Friday, July 30, 2010 - link

    I see the future! no local storage and not even processing. Praises to the light!
  • misterjohnnyt - Friday, July 30, 2010 - link

    Maybe we can build Positronic Brains with these...
  • Pinski - Friday, July 30, 2010 - link

    So, Intel is now at 50% of what another company(Infinera) is doing. And they're doing it with 10x10Gb wavelengths. With a plan to move to 5x100Gb wavelengths by 2012.

    I guess, it'd be nice to have this article maybe include a comparison to what other companies are already doing or planning on doing.
  • DanNeely - Friday, July 30, 2010 - link

    I guess, it'd be nice to have this comment maybe include a citation to what other companies are already doing or planning on doing.
  • Huron80 - Friday, July 30, 2010 - link

    Look at Infinera, aren't they a networking provider?

    As stated in the article, Intel is taking what is being done effectively in Fiber Optic network transmissions - the same ones that help connect the internet - and are making a miniature version that will work inside your computer - the connections on the motherboard to components.
  • has407 - Friday, July 30, 2010 - link

    Infinera's achievements are nothing to sneeze at, but they're quite different. AFAIK Infinera still uses a separate process for the optical components (the PIC, or "photonics IC"); they don't put everything on a single die. That's fine for certain markets, and they seem to have done well with Telecoms. However, for large scale low cost production; you're very unlikely to see it replace inter-chip or intra-chip interconnects in volume markets.

    Luxera is probably closer to what Intel is showing. They partnered with Freescale to integrate the photonics and electronics fab in CMOS last year to allow "low cost high volume" production"; exactly what may come from that is TBD. While we haven't seen much, Luxera's Blazar "Optical Active Cable" is a clue. But again, that's targeted at external inter-system connections; hard to tell if that's due to technology imitations or if Luxera is simply trying to generate revenue and build a market based on the technology is unclear.

    In short, it's all about being able to produce everything--optical and electronic components--using a common and low cost fab. In that, Luxera/Freescale and Intel seem to be in the lead, and I'd wager Intel is most likely to bring it mainstream sooner rather than later.

    This has the potential to significantly reduce the "interconnect tax" imposed by copper, which increases as speeds increase, and will soon be untenable. Intel said 20GHz was the ceiling for serial interconnects some years ago. (Teething problems related to power consumption with 10Gbe copper provide clues.) Even Inte's prediction is off by a factor of two, the end of that road is in sight.

    At a guess, we'll see photonics integration in high-margin parts with the first replacing QPI as the backbone interconnect in high-end systems. Or maybe blade systems with optical backplanes. Or...
  • Jaybus - Friday, July 30, 2010 - link

    Another thing to consider is that one reason for motherboards being as large as they are is that there must be hundreds of traces interconnecting components. Components have to be spread out to leave room for the traces. This technology will allow for one optical link to replace all of the myriad traces connecting the CPU to the northbridge, for example. Components can be closer together.

    Another factor is bus cards. Cards can be placed anywhere and connected by a single small fiber optic cable. Disk drives, etc. can be connected optically. This will substantially reduce the size of a full featured motherboard.
  • arnavvdesai - Friday, July 30, 2010 - link

    While I agree components can be closer. We also have to think about the issue of heat dissipation. If we start moving items closer then we might have to deal with other issues.
  • Jaybus - Monday, August 2, 2010 - link

    Certainly, cooling will force a minimum size, but there is no reason an ATX board could not be shrunk to ITX size, at least, yet have even more i/o capability than a current ATX board. The board components could be placed as closely as is feasible, given the heat dissipation required.
  • clarkn0va - Tuesday, August 3, 2010 - link

    Doesn't replacing electrons with photons reduce the amount of heat produced?
  • GullLars - Wednesday, August 4, 2010 - link

    Not necessarily, you still need energy converting the digital electronic signal to optical and back, it's just the transport that saves energy, and only as much as the resistance in the copper wire. For high frequency or long distance information transport, it can be noticeable power savings.

    Regarding cooling, one quick solution is to switch the default from air cooling to liquid cooling in higher power systems. By placing the radiator, pump, and reservoir outside the system, and exploiting the new available dimension inside the case, you can make much smaller (and quieter while better cooled) systems.
  • Ninhalem - Friday, July 30, 2010 - link

    The problem with a shrinking motherboard is that would leave too little room to put in cooling for the CPU and the memory. We're still a ways away from having CPUs running on light like the article said. I'm interested in not only shrinking the board a bit but having a different shape and orientation for the memory, video, audio card, and any other component you have in the case.

    Maybe have an upside down L shaped board where you can place the SATA II ports close to the hard drives and the optical drive. Maybe a start shaped board. Now that distance doesn't matter, the possibilities are endless. Very fascinating and exciting stuff.
  • darckhart - Friday, July 30, 2010 - link

    sure, but they haven't talked about heat constraints yet.
  • GullLars - Wednesday, August 4, 2010 - link

    Go liquid cooling, and you can get 500-2000W worth of power as densely packed as an ITX card. By using optical connections to build in 3D, you can have multiple planes of components.
  • nafhan - Friday, July 30, 2010 - link

    Another problem with motherboard design that this will help with is RF interferance.
  • nafhan - Friday, July 30, 2010 - link

    For anyone with a radio background, if the term WDM seems odd it's the term used by fiber optic people for FDM.
    Also, like to emphasize how big of an advantage that photonics/fiber optics have over normal electrical signals in motherboard design or any other application with a high frequency electrical signal. Replacing electrical signals with something that has essentially no radiated signal is huge. No interference is a very good thing.
    This stuff is just cool...
  • spunlex - Friday, July 30, 2010 - link

    <quote> it emits primarily phonons (lattice vibrations - and through a ton of hand waving and thermodynamic processes, heat), and very few electrons. </quote>

    I think this should read "and very few <b>photons</b>."

    Great article, it's good to see manufacturing of cheap quality lazers really took of in the past couple year. I am looking forward to many more interesting articles on photonics.
  • spunlex - Friday, July 30, 2010 - link

    I totally forgot these comments don't support formating, still pining for the edit feature.
  • Brian Klug - Friday, July 30, 2010 - link

    Oh wow good catch! That's what I get for writing this so late at night!

    -Brian
  • spunlex - Friday, July 30, 2010 - link

    I know what you mean, no where near as bad as some of the late night lab reports I have handed in.
  • iwodo - Friday, July 30, 2010 - link

    It mean we finally have VERY Good On broad Audio Chipset,
  • GDILord - Monday, August 2, 2010 - link

    Well, what exactly do broads have right now? :-)
  • cesthree - Saturday, July 31, 2010 - link

    When can we get this for an external GPU case?

    Have a case that only houses GPU's with their own PSU's and cooling solutions.

    Connect that case with your case via a Photon Link. I can't begin to list the advantages of a setup like this.

    I'm sure there is disadvantages. I know I would buy/make one if I could.
  • toktok - Sunday, August 1, 2010 - link

    Pick 'n' mix motherboards!
  • joshv - Monday, August 2, 2010 - link

    Wonder if different parts of a chip couldn't use optical interconnects to talk to other parts of the chip, basically in free space above the chip. Basically just point the detectors and emitters at each other. The nice thing is that the optical paths can intersect each other with no penalty. Electrical paths have to be routed around each other in the various layers of the chip.
  • Shadowmaster625 - Monday, August 2, 2010 - link

    Doesnt this make it possible to place 8GB of DDR3 onto a single chip? Just stack a bunch of dies and wire them all to a silicon photonic transmitter, then connect that directly to the cpu. Also, shouldnt this make it possible to stack all the NAND flash you'd ever need onto one die? And then SSDs can be made with one memory chip. Replace SATA with silicon photonics, and it should be possible to have a 100GB/sec SSD. In other words, there would be no need for RAM at all...
  • GullLars - Wednesday, August 4, 2010 - link

    There's a couple of fundamental problems with your last part.
    First is the method of read/write and R/W asymmetry of NAND.
    Second is latency. NAND is roughly 1000 times slower than DRAM.

    It would be great for scaling bandwidth though, but then you have the problem of data parallelism and transfer sizes...
    You would also need processing power for NAND upkeep (wear leveling, ECC, block management, channel management, etc), which would be a lot at 100GB/s.
    Today, the fastest NAND dies are around 100MB/s(+?) for reads, so you would need to manage 10.000 dies (which wouldn't fit one package BTW XD) for 100GB/s.
  • EddyKilowatt - Monday, August 2, 2010 - link

    ... whatever you do, don't dumb down AnandTech! This was perfect for a lot of geeky folks out here. Who else puts things like "Laser = gain medium + resonator + pump" in one non-stuffy, easily understood, non-patronizing paragraph?

    A few decades ago when I was a cub engineer, my mentor predicted almost exactly this as the logical evolution of inter-chip links. I hope he's still around to see this finally hit the market.

    How is Intel going to handle the I/O's to these optically-enabled chips? It's the only box on the "Siliconizing" viewgraph without a picture! Fiber is cheap but fiber termination is ex-pen-sive...
  • Cogman - Tuesday, August 3, 2010 - link

    Optics have far more benefits over just "really high speed" For starters, optical cables can be VERY close together without causing any sort of interference. This is a big plus for people like motherboard designers. Now, they could stick ram anywhere on the board (not just right next to the CPU) They can overlap connectors, do whatever. It makes a big difference.

    Besides basically having no crosstalk / interference problems, Optical transmission has the added benefit of being able to send multiple signals down the same line (we probably aren't to the point of being able to realize that.) It is entirely possible that we could get to the point of having two optics lines running from the ram, one for upstream, the other for down. Each having the benefit of being able to send full 256/512 (or higher) bits of information in a single transmission cycle. We send data serially now because of interference problems, with optics, that would no longer be an issue, parallel data transmission would easily be able to overtake its serial counterpart.

    Though, all the ultrahigh speed data transmission in the world doesn't mean a thing if the processor can't crunch the numbers fast enough.
  • has407 - Wednesday, August 4, 2010 - link

    > Now, they could stick ram anywhere on the board (not just right next to the CPU) They can overlap connectors, do whatever.

    There's still speed-of-light constraints. For the foreseeable future, the closer, the better (that's the primary reason for the CRAY-1 physical structure--to reduce interconnect distance, not because it looked cool).

    > Optical transmission has the added benefit of being able to send multiple signals down the same line (we probably aren't to the point of being able to realize that.)

    We've been there and done that for years. It's called WDM or DWDM. Virtually all fiber-based communication today multiplexes multiple optical wavelengths/signals across a single fiber. That's how we get much more bandwidth without having to lay new fiber.

    > We send data serially now because of interference problems, with optics, that would no longer be an issue, parallel data transmission would easily be able to overtake its serial counterpart.

    Ummm...no. Parallel interconnects, regardless of medium, suffer from skew and the need for deskew. That always has, and always will be, a problem for parallel; optical interconnects do not fundamentally change the equation (although they minimize some problems). Thus the desire to reduce the number of parallel interconnects by using a smaller number of faster serial interconnects. Optics offers the option of significantly increasing serial interconnect speed, thus reducing the need for parallel interconnects and their associated problems.

    I.e., the number of bits/sec that can be transferred over a single "wire" (or in this case fiber), minimizes the need for parallel interconnects. Maybe a few niche apps will need the augmented bandwidth that parallel can provide, but I have my doubts... Will WDM/DWDM scale fast enough? I bet it will scale faster/cheaper than parallel interconnects, at least for 99% of the market.
  • jpbattaille - Thursday, August 5, 2010 - link

    Totally agree with this comment - the article seems to make abstraction of these facts.

    Optical speed of light in fiber is 200 m/s, in vacuum 300 m/s.
  • jpbattaille - Thursday, August 5, 2010 - link

    Sorry, 20 cm /ns and 30 cm/ns
  • PlugAndTweak - Wednesday, August 4, 2010 - link

    This sounds very cool and promising on the paper.
    But what does it really mean to us as single end users?
    This isn't the components themselves we are talking about that will get silicon photonics, that was WAY ahead in the future.
    It's the interconnects.
    I bet that even if we would get all the interconnects in the mobo;s with Silicon Photonics we wouldn't notice any dramatic differences even in very intensive stuff like Videoediting, 3d-rendering, software synthesizers and plug-ins in DAW;s and so on.
    And would SSD;s REALLY be faster than any comparable technology today?
    Even theoretically?
    And take a look at tests where different memory speeds are done, there has always been minimal differences (often just 1-2 percent in average) when doing these tests. Increasing the bandwitdth in the interconnects won't have any dramatic difference in that aspect.
    Not for us single power users.
    For the big players with setups of loads of CPU;s and other components I could see the speed advantage.
    And there are the other mentioned advantages for the single power user, like less power consumed, possibly smaller motherboards (if they find inexpensive cooling solutions).
    Then there are all the advantages for the manufacturers with less interference and all the other stuff mentioned.
    I really think that the ones that will benefit most from this is the manufacturers of the motherboard, much more than us single users.
    Not until they get to the point of using it on the component level.
  • ghot - Thursday, August 5, 2010 - link

    "Blinded by the light...." I'm 53 years old and having been waiting for this since the 1st TI calculator hit the market in 1976. Unfortunately I probably won't live to see the 1st totally optical, consumer affordable PC.
    This doesn't prove 'Murphy's Law....it squares or even cubes it.

    The best part, isn't even the speed (although it will be incredible) but rather the fact that PCB printed circuits can be a long as the design requires. Imagine a motherboard 5 inches wide and 3-4 foot tall. Then a PC enclosure could look more like a 4 foot tall 6x6 inch speaker enclosure for example.
    Cooling needs will be greatly reduced as well. Just try to imagine the over clocking potential.

    I used to drool over the 1st commercially available digital watch...the Pulsar....didn't even have LED's...it used tiny Nixie Tubes.....and now this.

    When most of you are my age, you'll be complaining that your brand new Intel P4 (yes P4..maybe in YOUR lifetime, Intel will get back to simple naming schemes) is only capable of 4Thz on air. And, as every over clocker knows....increase the speeds in small increments...say .....10Ghz each time...lol

    I am so jealous :)
  • ziaullahk - Sunday, February 18, 2018 - link

    And to think about it, we just reached 100GBps according to https://compoundtek.com/news

Log in

Don't have an account? Sign up now