If current propagates at 2/3 lightspeed in copper why is adding optics such a big deal. It will only get you 33% fast performance at most right? Aren't the challenges of optics also a big hurdle like the refractive index changing with the temperature?
Adding optics to circuits
- Thread starter Stiganator
- Start date
Well good points.
The advantage isn't so much about propagation delay / speed as it may be about other factors.
Actually what the exact advantages are do depend on the implementation of the technology; it is certainly possible to devise hybrid optical ICs that are slower than and more expensive than just doing the similar circuit in pure silicon.
Now that they've got a few promising options to make monolithic or at least easily integrated light sources in the Silicon wafer manufacturing processing that's very helpful for cost / complexity of manufacture.
Optical is obviously nice for when you have to transmit signals into fiber optics to go long distances at potentially high speeds.
Optical might be nice for purely high speed reasons on short (intra-chip or inter-chip/intra-circuit board) interconnects, but the advances in multi-gigabit serial interconnects like HyperTransport, et. al. do cut away from the necessity to use optical methods.
Optical can be nice for sheer density of information since you can simultaneously have thousands of independent frequency division multiplexed channels on a single fiber, though the practicality of doing that to a high level of simultaneity within a short distance link is a bit questionable.
Optical does let you start to do things like quantum computing in certain ways for instance the recent articles on quantum teleportation of entangled atoms / electrons coupled by a fiber link, or generating encryption systems using entangled photon pairs, et. al.
Certainly it could reduce circuit design complexity and printed circuit board costs and chip packaging costs -- now you might have a CPU with a 64 bit data bus and 64 bit address bus and many such other pins. All those contacts that have to be routed from the die to the package, from the package to the PCB, and on 4 or 6 layers of PCB makes for a very complex routing problem since you can't cross the wires and building "overpasses" on another layer is expensive and takes a lot of space, etc. The mechanical solder joints are prone to failure and so on.
Now imagine you had some 3d chip packaging like a bunch of LEGOs, each one had a standardized optical interconnect on each of its six sides so it could talk to its neighbors if they were present. Then you wouldn't need a PCB, just a simple proximity based interconnection. Of course you could do the same with an electrical connector, but once you get more than a couple of dozen pins in any given connector the wiring starts to get inconvenient. So it really comes down to necessary bandwidth and cost of connectorization.
The refractive index of a waveguide and its environment are responsible for guiding the waves, yes, and if the index changes too much you could change the mode of propagation. It is mainly an issue if you needed say single mode propagation with very low loss, so you'd have to engineer the materials so that it is easy to lay down the layer thicknesses with sufficient quality that over the whole range of process variations, age, temperature, impurities, etc. it will do the job. I guess it's just an aspect of optical circuit design just as dealing with electrical noise is an aspect of electronic design.
The advantage isn't so much about propagation delay / speed as it may be about other factors.
Actually what the exact advantages are do depend on the implementation of the technology; it is certainly possible to devise hybrid optical ICs that are slower than and more expensive than just doing the similar circuit in pure silicon.
Now that they've got a few promising options to make monolithic or at least easily integrated light sources in the Silicon wafer manufacturing processing that's very helpful for cost / complexity of manufacture.
Optical is obviously nice for when you have to transmit signals into fiber optics to go long distances at potentially high speeds.
Optical might be nice for purely high speed reasons on short (intra-chip or inter-chip/intra-circuit board) interconnects, but the advances in multi-gigabit serial interconnects like HyperTransport, et. al. do cut away from the necessity to use optical methods.
Optical can be nice for sheer density of information since you can simultaneously have thousands of independent frequency division multiplexed channels on a single fiber, though the practicality of doing that to a high level of simultaneity within a short distance link is a bit questionable.
Optical does let you start to do things like quantum computing in certain ways for instance the recent articles on quantum teleportation of entangled atoms / electrons coupled by a fiber link, or generating encryption systems using entangled photon pairs, et. al.
Certainly it could reduce circuit design complexity and printed circuit board costs and chip packaging costs -- now you might have a CPU with a 64 bit data bus and 64 bit address bus and many such other pins. All those contacts that have to be routed from the die to the package, from the package to the PCB, and on 4 or 6 layers of PCB makes for a very complex routing problem since you can't cross the wires and building "overpasses" on another layer is expensive and takes a lot of space, etc. The mechanical solder joints are prone to failure and so on.
Now imagine you had some 3d chip packaging like a bunch of LEGOs, each one had a standardized optical interconnect on each of its six sides so it could talk to its neighbors if they were present. Then you wouldn't need a PCB, just a simple proximity based interconnection. Of course you could do the same with an electrical connector, but once you get more than a couple of dozen pins in any given connector the wiring starts to get inconvenient. So it really comes down to necessary bandwidth and cost of connectorization.
The refractive index of a waveguide and its environment are responsible for guiding the waves, yes, and if the index changes too much you could change the mode of propagation. It is mainly an issue if you needed say single mode propagation with very low loss, so you'd have to engineer the materials so that it is easy to lay down the layer thicknesses with sufficient quality that over the whole range of process variations, age, temperature, impurities, etc. it will do the job. I guess it's just an aspect of optical circuit design just as dealing with electrical noise is an aspect of electronic design.
Computational speed is limited by how quickly the metal logic gates can move, not by the speed of electrons or the speed of light. (We've been using the same electrons all these years.)
Further slowdowns come from a combination of heat (from electrical resistence) and signal leakage, and these are becoming more problematic every time we shrink transistors. Light doesn't generate heat, or leak, so generally speaking, any circuits we can convert to optics will run cooler and more reliably.
Further slowdowns come from a combination of heat (from electrical resistence) and signal leakage, and these are becoming more problematic every time we shrink transistors. Light doesn't generate heat, or leak, so generally speaking, any circuits we can convert to optics will run cooler and more reliably.
Optical-based computing would not, at least in theory, have the heat/thermal limitations that conductor-based circuitry. Of course you have to look at the total package, what are the light sources, any conversions, etc.
When talking about signal propagation within a CPU, it doesn't propagate anywhere close to that fast. A wave may propagate that fast, but the data won't. To explain, imagine a wire as a trough of water. If it has some water in it and you splash on one end, the ripples will travel quickly to the other end (let's say 2/3rds of c). However, that isn't how signaling actually happens inside a CPU - instead, you signal a zero by emptying the trough, and a 1 by filling it up all the way. Even though waves (ripples) travel very quickly, actually charging/discharging a wire (filling/emptying the trough) takes much longer than you'd expect based on wave-propagation speeds. Signaling with waves is much harder to do, and requires drastically larger devices to drive and receive the signals, and the wires have to meet certain requirements to propagate waves effectively.
Consider that people say you can't really get across a modern CPU in 1 clock cycle. Let's say a modern CPU is about 1cm long and runs at 3GHz, so 1cm takes at least 333ps which works out to 10% of the speed of light. Reaching c now sounds a lot more appealing, doesn't it?
Outside of an individual chip, at the board / rack / server-room level, I think signaling is generally done with waves, because the tradeoff makes sense (signal propagation speed vs. size and complexity of the driving/receiving logic).
There's a paper here that discusses using transmission line techniques within a CPU to talk to caches faster. Look at figure 3 to see how much bigger the wires are than normal wires in a chip - that means that any data coming from the cache is going to arrive with its bits very far apart, and that you can have very few of these wires on a chip just because they're so big.
It might not even come close to that. For example, fiber optic cabling used for networking operates at ~.66 c as well. Good coax (foam core, 84% velocity factor) can propagate signal faster then most fiber optic cabling (~66% velocity factor, as implemented).
FWIW
Scott
Modelworks
Lifer
- Feb 22, 2007
- 16,240
- 7
- 76
Another reason to use is optical is isolation.
Its the easiest way to connect two devices/circuits and have the two completely isolated from each other.
Its the easiest way to connect two devices/circuits and have the two completely isolated from each other.
Lower Power
-Imagine large server farms, power savings would be astronomical.
No Leakage (mostly CPU)
-A lot of the area of the core is there to protect against leakage ( eg Hi/Low K ) without any leakage you can drastically reduce the area of the CPU. ( Profit! )
-Also allows you to shrink transistors more easily. ( Again more profits )
-CPU speed is limited by the resistance(heat)/leakage the colder your chip runs the higher you can push the signal without leakage making you CPU become unstable. With little heat and no leakage we should be able to push CPU frequency's considerably.
Less Power, Smaller Footprint, Increased Performance what server company wouldn't love to buy thousands of those CPUs. This does not only affect CPUs though I just used that as an example but transfer data etc is faster and requires less power than over copper.
TuxDave
Lifer
- Oct 8, 2002
- 10,571
- 3
- 71
I'm not sure if I agree with any of your points. Layout density of metal interconnects and device densities of CMOS devices are very high and it's going to take a long tme to make optical drivers and optical interconnects to reach that level of density and "reduce your area of your CPU".
I think the biggest gain with optics is the lack of reluctance and electromagnetic field in the circuitry. That would be why I would try to make a circuit based on Optical instead of Electrical signals;the noise and filtering needed would be negligible in comparison to the electrical circuitry. It would make a whole lot of things so much easier, you wouldn't have to worry about running parallel lines causing unwanted capacitance, or other various EMI fields messing with your signals (although I think they would alter light as well, just not as much - not really an expert on that) You could make a circuit as large as you want, without having to worry about the EMI it will cause (because it won't cause any). So many design constraints would be thrown out the window if optical circuits were feasible. Of course, they would probably introduce new constraints that I have no experience in, so who knows if it owuld really be worth it?
degibson
Golden Member
- Mar 21, 2008
- 1,389
- 0
- 0
Another potential benefit to optics is the nature of optical propagation -- electrical signals in digital systems don't propagate in an on-off fashion, they are rather more like water slowly sloshing from one end of a tilted bathtub to another.
The ability to send data cross-chip without the need for 1) repeaters, 2) latches, 3) wide wires and 4) fast is really quite appealing indeed.
/nod to CTho9305 for citing Beckmann et. al. on Transmission LInes. Optics seem like a similar tradeoff.
The ability to send data cross-chip without the need for 1) repeaters, 2) latches, 3) wide wires and 4) fast is really quite appealing indeed.
/nod to CTho9305 for citing Beckmann et. al. on Transmission LInes. Optics seem like a similar tradeoff.
TRENDING THREADS
-
Discussion Zen 5 Speculation (EPYC Turin and Strix Point/Granite Ridge - Ryzen 9000)- Started by DisEnchantment
- Replies: 25K
-
Discussion Intel Meteor, Arrow, Lunar & Panther Lakes + WCL Discussion Threads
- Started by Tigerick
- Replies: 24K
-
Discussion Intel current and future Lakes & Rapids thread- Started by TheF34RChannel
- Replies: 23K
-
-
Facebook
Twitter