Slashdot Mirror
IBM Discovery May Lead To Exascale Supercomputers
alphadogg writes "IBM researchers have made a breakthrough in using pulses of light to accelerate data transfer between chips, something they say could boost the performance of supercomputers by more than a thousand times. The new technology, called CMOS Integrated Silicon Nanophotonics, integrates electrical and optical modules on a single piece of silicon, allowing electrical signals created at the transistor level to be converted into pulses of light that allow chips to communicate at faster speeds, said Will Green, silicon photonics research scientist at IBM. The technology could lead to massive advances in the power of supercomputers, according to IBM."
I hope they're also talking about standard GPU supercomputers (which are as we know pretty cheap at less than £100 for the low end!)
Why OpalCalc is the best Windows calc
Ah, just like the Ancients in Stargate. Now we only need to figure out the Zero-Point-Module.
Interesting, perhaps in our lifetime we'll see it make its way to the desktop... It'll be sorely needed by then to run Windows 2030 comfortably, until of course, the malware takes hold.
Just yesterday I was talking with someone about how IBM have devolved into patent trolls that do no worthwhile research. Have I been proven wrong, or is this just vaporware? Anyone with the knowledge to do so, please enlighten me!
weinersmith
IBM's press release is http://www-03.ibm.com/press/us/en/pressrelease/33115.wss
One interesting bit is that the new IBM technology can be produced on the front-end of a standard CMOS manufacturing line and requires no new or special tooling. With this approach, silicon transistors can share the same silicon layer with silicon nanophotonics devices. To make this approach possible, IBM researchers have developed a suite of integrated ultra-compact active and passive silicon nanophotonics devices that are all scaled down to the diffraction limit - the smallest size that dielectric optics can afford.
A whole dictionary full of perfectly good words and they have to make one up to mean “very large”...
Distributed Denial of APK: It takes 15 seconds to reply to him anonymously, but wastes tons of his time if we all do it.
I can play crysis.
...that the metal connections between individual components would not be fast enough.
I only wonder how long before this sort of technology makes its way to the consumer market, if only for show. Of course I can't see a use for an exascale databus on the mobo anytime soon.
the difference between this and Intel's technology, other than the obvious chip-to-chip vs machine-to-peripheral difference.
It's all variations on silicon (nano)photonics, right? The article says "Intel is also researching silicon nanophotonics at the silicon level, but has not yet demonstrated the integration of photonics with electronics"...but that makes me wonder what the big deal about Light Peak is, then... is the only difference the "nano"?
We have reached an informational threshold which can only be crossed by harnessing the speed of light directly. The quickest computations require the fastest possible particles moving along the shortest paths. Since the capability now exists to take our information directly from photons travelling molecular distances, the final act of the information revolution will soon be upon us.
-- Academician Prokhor Zakharov, "For I Have Tasted The Fruit"
Now I just need room temperature superconductors to build my gatling laser speeders.
Where we get a Positronic Brain from this.
I call it 'The Aristocrats'
... a Beowulf cluster of these!!!!
The interconnects are not the entire problem. Faster transmit helps, of course. But the information still has to come in from storage; it's still held in slow memory banks; it still has to propagate across the swarm. Software still has to be able to access that data in a way that makes sense and can scale to half a million nodes. Connectionless distributed computation is nontrivial, and while lower-latency intranode communication might get us the last 5% it won't get us the first 95%.
Did they publish a paper? The marketing talk is useless. I want to read what they really did. News websites should always link the Paper even if it's behind a paywall...
Optoelectronics really is the holy grail of computing. There's no cross talk problems, no magnetic fields to worry about, and you can multiplex the hell out of a communication link. The current record is 155 channels of 100 Gbit/s each. (!)
Now we can keep track of the deficit in real time.
He's sped up links between chips from something like one-third c to c.
Architecturally that reduces inter-chip latency by 66%, which does indeed open up a new overall speed range for applications that are bandwidth-limited by interconnects. But in no sense does it imply a 1000-fold increase in overall performance. It's only a 3X improvement in bandwidth of the physical layer of the interconnect to which the speedup applies.
It may allow architectures that pack in more computing units, since light beams don't interfere physically or electrically the way wires do. And light can carry multiple channels in the same beam if multiple frequency or phase or polarization accesses can be added. Those will further improve bandwidth and possibly allow a further increase in the number of computing units, which could help get to the 1000X number.
BTW, didn't Intel have an announcement on optical interconnects just a while ago? Yes. They did.
Are they implying that signals travel faster through a fiber optic cable than a copper cable? Or just that there is less interference between the lines.
I've abandoned my search for truth; now I'm just looking for some useful delusions.
Well, at least they are using their smarts to actually invent the things they claim instead of sitting on patents like some other companies. Now to remember the new password standard, minimum 90 characters.
...Of course I can't see a use for an exascale databus on the mobo anytime soon.
You'll need it to run Windows (insert release N+1 here).
Windows will finally be usable?
Please do not read this sig. Thank you.
... the Singularity must almost be upon us. I, for one, welcome our new supercomputing overlords!
(No it isn't, Ray Kurzweil is an idiot, and don't call me Shirley!)
FTA:
The photonics technology [will] help IBM to achieve its goal of building an exascale computer by 2020
So I guess IBM is in line with the International Technology Roadmap for Semiconductors.
There has been a lot of research done by the major players in the industry, individual components have been developped (light sources, couplers, phodetectors, optical waveguides, etc...) and IBM just showed they can produce them on-die with standard semiconductor production methods.
That's not the kind of breakthrough the article claims, it is usual incremental progress. And I am quite happy with that.
Kurzweil's not looking quite as crazy right now.
GPUs are indeed an inexpensive way to boost speed in some cases. But they have been rather oversold; while some specific types of problems benefit a lot from them, many problems do not.
Where do you get the idea that GPUs have been oversold? Is the loudest mouth breather in the room representative of the general consensus? One vain, overreaching guy from 1960 who had spent too many hours hunched over a keyboard predicts human level AI within the decade, and the entire endeavour is tainted forever? All to alleviate one slow news day?
2000 BC called, and wants their sampling procedure back.
Sixteen lanes of PCI-e V 3.0 has a architectural bandwidth of 16 GB/s and we're looking at about 4GB of local memory with very high bandwidth. The computational problems you can parcel up within these constraints is not small. This is before integrated GPUs becomes commonplace fabricated on the same die, with or without the IBM pixie dust.
For thirty years we enjoyed the regime that a rising tide lifts all boats. Everyone worked within a single dominant programming paradigm and (nearly) every program benefited from clock speed. Memory latency was drifting to the event horizon, but we mostly dealt with that. Until we hit the thermal bend.
If that is the standard of reference, *everything* from the last ten years was oversold.
There's a similar problem measuring the inflation rate. If you keep a fixed basket of goods, you get a consistent measure of inflation that becomes increasingly irrelevant. Back in 1970, they gave you an organic apple for the price of a regular apple. But aside from that, who wants the 1970 basket of goods, even at half it's original price? In that basket, an iPhone would cost you a trillion dollars, and take thirty years to deliver.
If you update the basket to account for the change in the kinds of goods available, you end up with a lower measure of inflation. When a Cray is small enough to fall into the toilet, and you don't even hear the splash, that's serious deflation. And what value do you assign to gene sequencing bird flu over the weekend? In your standard basket?
Exploiting the GPU is not a quick fix. There are immense transition costs, and it only applies to suitable computations. Tragic.
But think about the evolving basket of goods. Here's a good way to do it. Take every package in the R programming language and ask what level of acceleration is available from GPU computing, once the transition costs are paid in full, and then decide if the GPU was oversold or not. R covers econometrics, data mining, and computational biology to name just a few.
That basket works for me. Or do you think that statistical inference over massive data sets has no relevance to the next twenty years? Are you saying that massive stockpiles of data are oversold? Really? That's a brave position.
I recall serious observations circa 1990 from prominent economists that the PC showed up everywhere in the economic data except for productivity. Relative to the 1980 basket of productivity, there was merit in the argument. In 1980, you changed your font once every two years. To successfully fritter, you had to play Pac-Man on CGA.
I get miffed when the standard of "oversold" is denominated in News at Eleven squeaky-wheel gratification reflex, while reeducating the entire white collar work force to a whole new way of doing things is taken for granted.
One man's oversold is another man's relentless progress.
When are we going to start rating software in cycles/keystroke or any kind of metric? Why is it always the job of hardware designers to rescue your sorry, undisciplined asses?
I remember beading an article in the IEEE Spectrum called: "Reaching for the megaflop" in the nineteen-seventies.
I was working for CDC building power supplies (at their facility in Dorval, PQ, Canada) and keeping up with technology.
Exaflop computing is just blowing me away...
MSBPodcast.com The opinions expressed here are my own. If you don't like 'em... Think up your own stuff.
More information available at source.
They replaced metal interconnects with light? Yawn. Wake me up when we hit wormholes.
I remember when IBM's future PowerPCs would be a gajillion times faster than anything from Intel and then Apple would finally rule the desktop. ...
Oh wait.