Slashdot Mirror
Holy Grail "Opt-Chip" - 100GB/sec?
silicon_synapse writes, "ZDNET has a story about the new Opto-chip which can supposedly transfer data at 100GB per second. Yes, gigaBYTES. A two-hour digital movie could download in 1/20th of a second. The only problem is making the rest of the computer fast enough to take advantage of it. " The researchers are being published today in Science magazine and claim that the U.S. military will be using this as early as next summer. However, I think this is going to be another case of wait-and-see - the technology sounds a little too good - "spray on" application and such.
I'll preface my remarks by saying that I haven't read the Science article and the ZDnet article seems pretty short on scientific details. That said, this is one of my research topics, so I know a little bit about the area.
First off, I've seen some questions about the quote of "spraying" onto a chip. There are a variety of techniques, but I'm guessing they're talking about spin-coating or CVD (Chemical Vapor Deposition), which are both used routinely in manufacturing.
Secondly, these electro-optic devices use "second order nonlinear optics" (for all you physics geeks). Basically, people have been using crystalline modulators like lithium niobate for years, but they're very expensive and hard to make. So most of the research in the area has gone into making organic/polymer/self-assembled modulators. The idea is that you encase your chromophore molecule (the "active ingredient") in a polymer or other strong-film environment. Then you use this film in a waveguide and use it like a switch. The best mental picture would be a railroad switch--the electrical signal switches optical tracks for the optical beam.
Without reading the published results, it's hard to know if this is really a breakthrough. My questions would be whether it's actually a new chromophore that's giving better results, a better preparation method, or something else. It sounds like they're making some change to the preparation of polymer devices, which are behind the self-assembled films many labs are making now.
Suffice to say, the *real* revolution will come if anyone can get a usable third order NLO device. This would allow optical-optical switching.
-Geoff
Stupid question time - what is the maximum switching frequency of a plain old phototransistor?
Exotic technologies are very neat, but I'm wondering if more conventional technologies might already work.
If phototransistor switching speed is comparable to ordinary transistor switching speed, you could probably build an optical transciever more cheaply by using closely packed frequency channels with bandwidth comparable to the switching speed, and a prism or diffraction grating to split them for parallel reading.
I think you are missing the point here. The Bell Labs benchmark was done at 40gbs a channel. This story is talking about 800gbs on one stream of light or channel. (Not that I think /. is very current on its news either...)
"The experimental GigaChannel Ethernet multiplexer combines up to eight independent gigabit Ethernet signals into a single 10 Gb/s signal
stream."
It sounds dopey, but it actually makes sense - what they are saying is that they can fit 8 Gigabit ethernet channels into a single OC-192 carrier (OC-192 is an industry standard, ~10Gbps SONET optical data rate) They can't fit 10, because there is overhead in any SONET stream, and they'd need extra overhead to split out the 1Gbps channels from one another. It seems like they ought to be able to fit 9 channels in there, if they really wanted to.
A thought just occurred to me - they may not be TDM muxing the 8 signals at all, but rather they are saying that they can cram 8 1Gbps carrier signals into the same frequency range that would normally carry a single OC-192 carrier. This would make it easier (read: cheaper) to split out one channel to drop it out at it's destination without having to have expensive 10Gbps/1Gbps mux hardware at each terminus, and it is consistent with them needing to have guard bands [dead frequencies between the carriers so one signal doesn't stomp on the one next door] between the 8 carriers. The more I think about it, the more I'm convinced this must be what they are doing - the other way would be *way* too expensive.
So, loopy as it sounds, fitting 8 Gigabit Ethernet channels into the 'space' of a 10Gbps optical channel makes perfect sense when taken in context.
I hate it when people talk about discovering more bandwidth that we can use ( or process ). When has this _ever_ been a problem in our history. My immediate reaction to this is that if the MPAA were to read that article, they'd be needlessly ultra paranoid about DeCSS. I say this because you and I will NOT be able to practically download entire DVD's any time soon. In order for us to have 100G / s bandwidth right now with that technology, it would have to be an isolated point to point network. Which means that you will probably be very familiar with the other end of the connection. The fear of DVD copying usually involves complete access to the entire web and a random end point for transmittion. The statistical likelihood of a desired DVD being on your other end are rather slim.
Another use for this 100G network would be in an office situation but again, unless you've got a point to point star network, you're multiplexing someone's data which will reduce your bandwidth ( to well within a computer's processing capability ). Still, it's a hell of a lot more than what we have now.. But again, live video feeds are from static, known points ( mainly within your building, or between a finite number of known buildings ).
So then, let's apply this technology to general publicly available internet connections. What do we have.. Raw bandwidth that will be soaking up by Linux distribution downloads, pr0n and general web-site traffic. Meaning if you put more bandwidth up, then the population will increase the volume of it's downloads to fill it. It's a basic trend that I'd be hard pressed to not call fact. And lets put this into perspective.. These 100G connections out on the internet are hardly going to be noticed over the existing high bandwidth lines where the routers are the slow point. Yes you can put high perf routers, and yes you'll eventually be able to maximally traffic this data, but your 56K modem or 1Mps cable network is not going to take advantage of it. And I'm doubting that we'll see home connections any time soon. Buisnesses that can afford such a connection are probably going to be saturating it. This is because it's probably not going to be cheap, and a business doesn't usually indulge in a technology for economic efficiency reasons.
Will this help? Of course. Will it be great? Hell yeah. Will you realize any benifit? No. Because it's like a savings account interest rate when there's inflation. You may be getting 10% interest in your savings account, but if inflation is 15%, that doesn't get you anything more tomorrow than you have today ( you'll just be hurting less than if you only got 2% interest ).
The biggest threat I see to the internet is video feeds ( hence my focus in this article ). If the public sees high bandwidth, they typically chant video, which, in my mind, if it ever comes to fruition then imagine the effect of thousands of homes leaving their internet TV connection on all night ( like we do our internet radio here at work ). This is just a rant, but it reflects my current paranoia about public bandwidth.
-Michael
-Michael
Now now now, don't forget yourself. If you realy wanted to increase bandwidth, on a board, you'd either need fiber chanels or massively wide busses. Both of which are rather expensive, not just for the motherboard, but for the periferal manufacturers. The whole driving force in the PC industry has been supply and demand. IBM pushed MCA back in the 386 days ( absolutely better than ISA ( and probably even better than EISA ) ), but obviously the world didn't stampeed to this technology ( just like they don't stampeed to Alpha's apparent superiority, even though they can technically still run the same programs ). Granted IBM was proprietary. But the point is that the PC industry _can't_ just supe up their systems, because that costs money, and S & D requires an equilibrium for profit maximization ( and in the PC world, that just about breaks you even ).
There's also the case of compatibility. Even if you could produce as superior device in both performance _and_ price, you run the risk of lack of compatibility and thus you can only play a nitch and thus S&D kills you again.
The reason the PC industry has been technologically advancing so quickly is because there has been competition for maximally compatible components that simply run faster. ( Plus a segmented market with some willing to pay premium, and a large majority demanding the cheapest ). If we didn't have that diverse market, we'd still be running a 486 class machine today. ( And Alpha's wouldn't be windows 3.5 compatible )
-Michael
-Michael
- more pins (# bits/clock)
- higher clock rates (greater bits/time)
- move the bus on chip meaning you can't just plug in a new card
The first two mean more power - power being (very) roughly proportional to the number of pins and how fast they are waggling - and as a result hotter chips. The more pins solution breaks down pretty fast - you can double the bandwidth by doubling the bus width only so many times before running into packaging problems - remember at high frequencies you need 1 power/ground pin for every 3-4 signal pins - also plug-in card with >64 data pins are probably impractical Bumping the clock rate while keeping the bus narrow seems to be the way some parts of the industry's going (1394, RamBus, the new fast USB etc).For most of the PC space the third option is probably going to be what you see - more integration - buses going away or being pushed on-chip meaning that the chances for plug-in high bandwidth goodies are virtually non-existant - instead you get what was chosen for you by the person who chose the chip when they put the motherboard together.
Anyway the thing to remember TANSTAAFL - everything is a compromise.
... download a Jon Katz article without having to go for a cup of tea?
But seriously, this sounds like a great technology, and one needed to implement the "Internet of the Future", whatever that may be, put it is only one technology out of a host which are required. Sure, in the short term this will give rise to improvements in data transmission, but until a series of other breakthroughs are made this won't reveal its true potential.
So yeah, 100Gb/second is possible, but not for quite a while yet.
Some of the cool stuff some researchers are doing is integrating a laser onto a normal ASIC....
[...]
Now all we need is a way of producing RAM and peripherals that keep match with the speed....
For the RAM, at least, the answer is straightforward. Keep latency at its current range, but _heavily_ interleave RAM both on a bank level and a chip level. You now have RAM that can get 100 cache row requests and service all of them with a batch latency of 7 ns (or 5 ns or [etc]).
This would let you, say, put 8 or 16 cores on a die without worrying about cache misses slowing you down (as long as you have a deep miss buffer).
This would also be useful for transferring vast amounts of data with good locality in a known pattern (for instance, triangle or texture data) from RAM to a peripheral.
This is probably what busses will look like in a decade or two, as it's much easier to eliminate cross-talk and interference on an optical bus than on an electrical one.
Here is a research that is done at Lucent Technologies:
Instead of switching from optical wave to an electrical charge they use optical repeaters with mirrors and optical amplifiers.
"The DWDM-ready GigaChannel has been demonstrated over 40 kilometers of standard single-mode fiber using WaveStar MetroPoint and also over Lucent's flagship long-reach product, the Wavestar OLS 400G, using multiple 80-kilometer fiber spans with online erbium-doped optical amplifiers and dispersion compensation."
However it's only 10GB/s. Maybe they'll learn to do better than that.
"The experimental GigaChannel Ethernet multiplexer combines up to eight independent gigabit Ethernet signals into a single 10 Gb/s signal stream, enabling switches, routers and servers to connect at 10 Gb/s in native Ethernet format without the need for protocol conversion. The prototype complies with today's IEEE Gigabit Ethernet standard."
You can't handle the truth.
IO performance has always been a problem with PC's. We've had PC's around for how long... and all we have to show for it is AGP 4X????
While CPU horsepower has been following Moore's law pretty well, the PC world has lagged behind in terms of bus bandwidth. "100GB/sec" peripherals are useless when your bus runs at 133Mhz.
Let's start pushing chipset and memory manufacturers to start putting out faster busses and memory subsystems, and then PC's will finally begin to approach supercomputer-level performance.
________________________________
Terabit and faster networking isn't totally cutting edge anymore. Lucent is talking about sending many terabits per second over a single fiber.
What is interesting is the ability to process packets at that speed. This chip is critical in converting that optical stream into an electronic stream. The other part is a CPU or multi-CPU architecture to process the data. I'm sure Cisco is very interested in this.
So with Lucent figuring out how to send multiple terabits per second over a single fiber, this company able to convert those signals into electronic form, and hopefully soon Cisco being able to process and route data at those speeds, we'll soon be able to forget about bandwidth issues on the Internet. Or to be more precise, the bandwidth issues will become almost entirely limited to the link between consumers and their ISPs.