Great report on Oracle's progress in silicon photonics. Intel claims to be ahead of the pack, and has the systems to prove it, but as noted in Rick's story its still early in the game and Oracle has the advantage of its big installed base hungry for the affordable increased bandwidth.
Intel has been talking about its good research for a decade. Krishnamoorthy said Oracle/Sun has been working as long but saying less. Rather than in house, Oracle is doing more with partners--startups Kotura and Luxtera among them. The sense is everyone is at about the same stage of readiness for a ~2017 market.
What is the main distinction of photonics versus, say, using light in fiber communications?
Connectivity! Cheap, fast, powerful, and energetically low-cost flows of humongous numbers of bits between chips.
Imaging a gigantic server farm with only a couple of traditional analog phone lines running into it. You have enormous power, but no way to give it the data it needs to use that power, and no way to get the result back out.
Sadly, that's pretty much the situation for most modern chips: They are IO strangled, forced to live with data rates in and out that are far less than what they could handle if given the chance. It's like having a cheetah for a pet and keeping it in a room that only measures 5 meters by 5 meters.
The reason for this lies deep into physics: We use electrons to communicate with chips, and electrons are fermions, which just means they get very ornery about being pushed too close together. So to use them, you have to create separate paths for them, and also keep them separated in time. It's like herding cats: It only works if you define very clear paths that even an ornery cat has to stick to get anywhere.
Photons in contrast are bosons, which just means they have zero problems with stepping all over each other. So for example, in a cross-connect, photons can pass right through each, and do it in huge quantities, without any problem. You can see the same effect when you cross the beams of two lasers or flashlights without any problem (well, unless you are in a Ghost Busters movie). For contrast, try crossing the fermion-based water "beams" coming out of two garden hoses and watch the splatterific results. Ditto for electrons that are not kept carefully guided at all times.
Finally, photons can cross empty space, and have no annoying charges that must be cancelled out within a circuit. That makes them closest thing available to pure transfers of information.
Put all of that together, and what it means is that photonic lattices can unleash the real power of all those little chips, the ones that are pacing their silicon cages in abject frustration. Everything starts to move a lot faster, and does so using far less energy.
The two slides on Page 4 of the article nicely show the cost implications of letting all those little silicon cheetahs run free on a photonic lattice: A 6.5 times reduction in power costs, not bad! I suspect there would also be a huge reduction in the footprint for a given level of capabilities, so we are talking about *major* expansion of money-generating capacities for legacy data centers that are stuck with fixed power and cooling capacities.
So, as someone else noted here, it's great to see the old idea of photonic interconnects reaching a point where real commercial systems may be just a few years away. Success with photonic interconnects would obsolete traditional data center architectures and eventually take them over.
Whether Oracle will succeed in being the one to dominate this incipient market is yet to be seen, but clearly they are doing some great work. But my strong suspicion is that other groups are also doing great work, and some of them may not want to advertise what they are doing quite yet. That's because the market potential for this market so high for whoever gets it right first: Reinventing and re-equipping legacy and new data centers globally? Uh... wow? (A caveat: The second-starter-wins rule may apply here. Surprisingly often in IT, the first big player makes all the mistakes, while the second big player starts clean and ends up outpacing the founder.)
Good post. All comes down to Bose-Einstein vs Fermi-Diraq statistics. Ironically, what makes photons so good for carrying information over distance makes them so bad for processing information due to the lack of interaction. We will still need electrons for a while longer ):
Thanks for this great information. I am truly enlightened now on this topic.
In the early days of microprocessors , the processor speed used to be a limitation to handle the real time situations so much so that we had to count those CPU cycles while writing the I/O handling interrupts. Now it seems the other way round .
@Prabhakar: Yes, the Oracle presnetation did a great job putting into context expected advances in CPUs and memory that would leave interconnect as the next bottleneck in the near future and silicon photonics as the best route to busting thru it.
This is not as simple as it sounds...you can't put silicon photonics transceiver for every IO signal...you can (maybe) do it in few selected places...but again where is teh source of the optical clock coming from??? I doubt that 2017 will see commercial deployment, the technology is not ready yet...Kris
The key technical probolem for years have been building a laser on silicon die...without it it is hard to talk about silicon photonics...by the sound of teh article this key problem has not been solved yet
Most of the plastic waveguide components are easily integratable i.e. receivers/transmitters, etc. Indeed that's the attraction. Couplers can reside in the waveguide itself or the connector or the circuit board. Of course complete integration may not be possible - at least for some time, but same with Si photonics.
it is ture, lasers cost are 10 times of silicon photonic chips. evanscent coupling based laser has been demonstrated and they claimed that the cost is much lower than other laser vendor, but the overall power cost still high (Ith is high). Fujitsu is using butt coupling for laser, but the packaging cost may high, the best way is to have a good grating coupler to couple the light into silicon photonic chip vertcially.
It should be no problem to form fiber to waveguide automation alignment as long as you can have a good spot size converter to expand the mode to match to the single mode fiber. 1um automation alignment accuracy should be ok to achieve.