[HN Gopher] Intel: Advances in silicon photonics can break the I...
___________________________________________________________________
Intel: Advances in silicon photonics can break the I/O "power wall"
Author : alexrustic
Score : 56 points
Date : 2021-04-12 13:37 UTC (9 hours ago)
(HTM) web link (venturebeat.com)
(TXT) w3m dump (venturebeat.com)
| bullen wrote:
| Copper allready transports electromagnetic waves at the speed of
| light. You should probably view this as a desperate attempt to
| keep the status quo of eternal growth.
|
| The real reason for all the news you see today, and that you will
| see until the end of times is caused by peak dead trees.
|
| Or atleast until the sun has powered enough trees, which BTW is
| the best solar panel AND battery; to make the same amount of
| coal, oil and gas we had in the early 1900.
| jcranmer wrote:
| The speed of light in copper cable is _not_ the same as the
| speed of light in a vacuum. Light will propagate faster through
| optical cables than it will in copper cable.
| akvadrako wrote:
| They are almost the same. Speed of electrical signals in
| copper cable and speed of light in optical cables is about
| 2.0e8 m/s.
|
| Though you are technically correct that _Light will propagate
| faster through optical cables than it will in copper cable._
| annoyingnoob wrote:
| While this could allow you to put more distance between
| components, I'm not convinced that this is a good thing in
| practice.
| ksec wrote:
| Intel has this vision for a very long time with their Silicon
| Photonics in DataCenter. Where you decouple the CPU, Memory and
| Storage into separate server unit or even _racks_. Essentially
| allowing Datacenter as Computer and your Cloud services to spin
| up / fine grain your CPU / Memory / Storage on Demand.
|
| Still very far from reality though.
| FourHand451 wrote:
| Why do you think it might not be a good thing in practice?
| annoyingnoob wrote:
| > That could look like pools of compute, memory, and
| peripheral functionality distributed throughout the system;
| connected over long distances with optical links, software-
| defined infrastructure, and high-speed networking.
|
| Latency comes to mind. But also how many people does it take
| to troubleshoot a system where the compute, memory, and
| storage are all in different physical locations?
|
| Building a distributed system seems fragile.
| sumtechguy wrote:
| Very much so. It also creates some interesting
| possibilities. Such as a memory unit that can be shared
| between a few computers and then reconfigured as needed.
| This would be useful in such things as VMs and docker units
| and dynamic scaling. They probably would not be located in
| different buildings. But probably in the same rack or
| adjacent. You can see some of this sort of thing with
| software defined networks and storage pools.
| annoyingnoob wrote:
| Modern approach to mainframe-style computers, I guess.
| Where everything is a shareable resource and physical
| location is abstracted away.
| deburo wrote:
| Unrelated to photonics, but this:
|
| >memory unit that can be shared between a few computers
| and then reconfigured as needed
|
| It reminds me of reading recently about Compute Express
| Link, when Micron announced they were moving away from 3D
| Xpoint and would focus on CXL.
|
| https://blocksandfiles.com/2021/03/25/cxl-and-the-
| developing...
| sumtechguy wrote:
| Yeah that is an interesting gamble on the part of Micron
| (not a bad one at this point). In they are betting xpoint
| will not deliver the price points needed to compete with
| DRAM, might as well put a node in the same place with
| similar latency as xpoint. Intel and Micron were seeing
| xpoint as a layer in memory. With DRAM being a L4 cache
| to the data (or some mix). The problem is keeping xpoint
| prices bellow DRAM yet high enough to keep the thing
| profitable. Fiberoptic links between the two nodes at
| this point would probably really complicate things (at
| this point) as the tech just is not there yet. But if
| they can get it to work the latency should overcome some
| of the issues of a secondary node holding shared memory.
|
| All of that because Intel is having trouble moving to
| 11nm. All of their features were tied to that node
| process (tick tock). It is why we are seeing Intel back
| features from the 11nm node back up to 14. As their
| competitors are not sitting on 2017 tech and milking it.
| They are moving to the new stuff. xPoint was interesting
| 5 years ago. When they could make a 128GB stick for half
| the price of DRAM one. But they did not have the pins on
| the chip to support the larger workloads. DRAM caught up
| in size to overcome it. They have them now. But now that
| config is not as interesting.
| lazide wrote:
| For voltage transient (aka surge) protection this is essential
| - there is nothing that can reliably give you protection except
| for insulation and lots of it (and space/gaps with no
| conductive paths is a necessary part of that)
| fanf2 wrote:
| Who remembers Intel Light Peak from 2009? They switched from
| silicon photonics to copper and it became Thunderbolt in 2011.
| https://en.wikipedia.org/wiki/Thunderbolt_(interface)#Introd...
| m463 wrote:
| I think displayport does optical right now. all the dp 8k
| cables are fiber.
| dogma1138 wrote:
| There are some active cables for DisplayPort that are optical
| the port itself is still electrical it's basically like a
| fiber SPF module.
|
| The benefit is that you get a long, thin and flexible cable,
| the downside is that these cost and arm and a leg (PS400 in
| the UK for 15m cables).
| CoastalCoder wrote:
| I'm curious about the relative _latency_ of photonics vs.
| electronics, when connecting e.g. two chiplets in the same
| package.
|
| My (possibly wrong) understanding is that the E-O and O-E
| conversion adds a non-trivial amount of latency in a situation
| like that.
| jl2718 wrote:
| You can calculate it as the RC constant of the PN junction. The
| issue is that the photon wavelength makes a fairly large area.
| This may be solvable with plasmonics.
| adgjlsfhk1 wrote:
| Has there been much research on making cpus that are fully
| light based? It seems like if that were possible, it would make
| a lot of this much easier.
| N1H1L wrote:
| I think one of the problems would be miniaturization. Visible
| light is hundreds of nanometers, and X-ray photonics are
| still in it's infancy
| entropicdrifter wrote:
| You aren't limited by the wavelength in terms of pure
| transmission and reception. Bass frequency sound waves are
| multiple feet long, yet earbuds transfer those frequencies
| just fine.
| derefr wrote:
| A wave of sound is a wave of _pressure_ -- i.e. density.
| Sound waves are one-dimensional. You can have a material
| one atom thick in a vacuum, and it'll transmit sound just
| fine, as long as that material has some elasticity to it.
|
| Light waves, on the other hand, are two- [linearly-
| polarized] or three- [circularly-polarized] -dimensional.
| They're heading in one direction, and also wiggling side-
| to-side as they go.
|
| If light didn't have a "thickness" -- that is, if light
| could "sneak through" gaps narrower than its "diameter"
| -- then 1. "Optical" fibre would be capable of acting as
| a waveguide for all frequencies, not just optical ones;
| and 2. the Faraday cage around a microwave wouldn't
| prevent the microwaves from escaping.
|
| (You might think we'd also get universal antennas out of
| that, but no, antennas are a quantum thing--it's not
| about squeezing through a gap, but about exciting an AC
| electric field in the material, so letting the _whole_
| wavelength hit the antenna actually causes the induced
| field to destructively interfere with itself. Think of an
| antenna as absorbing the wave "head-on", rather than
| "side-on." Your antenna needs to catch _half_ the wave --
| node to anode to node -- to get the highest-power signal
| out. But you don't miss the wave entirely if you catch
| less. You just get a less powerful signal -- it's still
| possible to do useful work [like power an optical
| receiver, or decode an AM radio signal] with just a 1
| /Nth slice of each cycle of the wave.)
| CoastalCoder wrote:
| Lightmatter [0] is doing something along those lines. But
| IIUC they're only attempting optical _computing_ for certain
| mathematical ops (FMA and /or convolution?)
|
| [0] https://lightmatter.co/
| mmmBacon wrote:
| Matrix multiplication. They use a cascaded Mach-Zhender
| MEMs array as their systolic matrix multiplier. These types
| of schemes are limited to the amount of loss that they
| have. Additionally since these are analog components they
| have a certain physical size which is difficult to reduce.
| In my view this type of architecture does not scale very
| well.
| magicalhippo wrote:
| AFAIK yes. Of course you need some actual non-light hardware
| to guide the light and such though.
|
| https://en.wikipedia.org/wiki/Optical_computing
|
| https://www.eetimes.com/how-does-optical-computing-work/
| moistbar wrote:
| I remember hearing about one back in the early 2000's, but
| I'm unable to find the article I read way back then. I recall
| it making some extremely lofty claims regarding clock speed,
| somewhere on the order of 300 GHz IIRC, but it never saw the
| inside of a production machine as far as I know.
| musingsole wrote:
| Most projects I've read focus on specific components (mainly
| ALUs). I'm aware of a photonics flip-flop equivalent design,
| but I'm unaware of equivalents for dense memory. I imagine
| DRAM would remain mundane electronics even if the CPU were
| fully migrated to the tech.
| amelius wrote:
| I guess the main problem is that Maxwell's equations are
| linear, and thus you need some optical non-linear effect to
| make a logic gate. And you need some way to prevent
| attenuation of your signals.
| mmmBacon wrote:
| The latency at this scale of the device is largely irrelevant
| as it would be dominated by FEC. The PN junctions themselves
| have BW of about 25GHz, well above anything compute is doing.
| musingsole wrote:
| Do photonics based components have similar potential for complex
| calculations as quantum computing?
|
| We use binary because it was an improvement on analog, and it
| beat out ternary or higher orders. However, light can have a
| number of effectively binary attributes. I wonder if you could
| use polarity in addition to the light's presence for smuggling
| more data into a "bit"
| mmmBacon wrote:
| In theory yes. For example, certain optical elements perform
| Fourier transforms. However optical computing as we know it
| today is really a type of analog computer. Today's optical
| devices are relatively large compared to transistors so
| difficult to make an optical computer scale very well. Also
| given that each optical computing stage involves optical power
| loss, there's a limit to how many stages one can cascade before
| having to regenerate via an OEO conversion.
| pjc50 wrote:
| > the company hopes to scale its silicon photonics platform up to
| 1 Tb/s per fiber at 1pJ of energy consumed per bit, reaching
| distances of up to 1 km.
|
| Astonishing numbers. 1pJ is tiny. And yet that's still cooking
| away at a whole watt of TX power.
| ksec wrote:
| Yes. I wonder if this will be _the_ future of Intel. They seems
| to be quite far ahead in Silicon Photonics than every one else.
| And I dont see an end for our need for more bandwidth in the
| near future, at least not in the next 10 years.
| cogman10 wrote:
| Bandwidth isn't the reason photonics matter (IMO). The reason
| they matter is because they eliminate cross talk. That can
| mean more densely packed transistors and less data lines.
| rektide wrote:
| if we believe their press releases, they seem like they've
| been in the lead for the last 10 years[1].
|
| they had one of their very first demonstrations a year
| ago[2].
|
| it's promising as heck. and I dont mind that it's taking a
| while. but it is also an area I have, after the many years,
| learned to keep tempered expectations on.
|
| [1] https://arstechnica.com/gadgets/2010/07/the-future-of-
| electr...
|
| [2] https://www.servethehome.com/hands-on-with-the-intel-co-
| pack...
| mmmBacon wrote:
| Many others have been shipping SiP in higher density for
| much longer. Acacia and Luxtera (both acquired by Cisco)
| for example.
| oneplane wrote:
| This is where they can start thinking about 'optronics' as a
| marketing name, isn't it.
| neogodless wrote:
| No, this branch of semiconductor engineering is un-optane-able.
___________________________________________________________________
(page generated 2021-04-12 23:03 UTC)