facebook rss twitter

Silicon photonics' terabit bandwidth makes it the interconnect of the future

by Tarinder Sandhu on 18 September 2007, 06:53

Tags: Intel (NASDAQ:INTC)

Quick Link: HEXUS.net/qajuo

Add to My Vault: x

Silicon photonics

The rise of multi-core processors causes as many headaches as benefits in the design stage. How do you keep the cores all talking in a speedy fashion, how to hook-up multi-socketed systems with one another? How, in essence, do you ensure that bandwidth isn't the obvious limiting factor?

Dr Mario Paniccia announced that Intel has architected the world's best-performing silicon/germanium photo-detector for ultra-fast bandwidth. Thinking about massively multi-core processors, silicon photonics will help scale interconnects between processors, both in terms of on-die communication and inter-machine throughput.

Getting back to basics, photonics is the technology of emission, transmission, control and detection of light. At the moment, most photonic devices utilise highly exotic materials to achieve high-speed transmission. Such technology, Mario noted, is currently utilised in the telecoms industry at, obviously, high cost. Successfully basing photonics in ubiquitous (low-cost) silicon should allow it to become a commercial success in future processor architectures.

Now, Mario enthused with much gusto, Intel is internally demonstrating silicon laser modulars operating at 40GB/s - near the speed of the aforementioned exotic modulars that cost an order of magnitude greater. The inherent beauty of the silicon laser modular is that you can team them up to increase bandwidth, such that 25 can be combined to achieve a potential throughput of 1TB/s.

Germanium is used to increase the wavelength detection, as silicon is limited to around 1 micron, and adding germanium increases the wavelength to around 1.6 microns, but germanium is 4 per cent larger than silicon. Intel, then, requires a strain-balanced lattice, and Dr. Paniccia's team has been working on increasing both bandwidth and efficiency, for their photo detector, over the past 5 years.

Tying this together with tera-scale computing, fast interconnects - for inter-CPU communication and memory-to-CPU transmission - will be required in many-core processors - and high bandwidth is an obvious requirement. Further, linking processors via a high-speed interconnect, from rack-to-rack and board-to-board, will also becoming vitally important Here, Mario asserts, is where a small-sized ASIC, with multiple optical modulators, will come to the fore. To put this into some kind of context, the world's fastest interconnect speed record is held by NTT, with a 25TB/s throughput. We note that copper, as an interconnect, currently scales to 40GB/s and is cheaper to produce.

Mario reckons that the technology should see commercial application by the end of the decade - Larrabee and silicon photonics is a no-no, then.


HEXUS Forums :: 2 Comments

Login with Forum Account

Don't have an account? Register today!
Posted by awm - Tue 18 Sep 2007 07:10
Cool stuff, although reality could provide a cruel damper when it actually goes to a real fab on silicon intended for real use. I'm sure they'll figure it out eventually although were other funny stuff like quantum computing will be then is a big question.
Posted by roddines - Tue 18 Sep 2007 07:22
I saw this about 2 months ago somewhere on the net. Here is the best place to find out about it.
http://blogs.intel.com/research/2007/07/40g_modulator.html

I think Intel would pretty much know they can make it before they would publicly declare it. So I think we will “see” it soon enough (although someone in the blog asks the question about an equally high performing optical detector/demodulator which seems to be unanswerred). It will at least intially be expensive and probably relegated to the supercomputing arena eventually we may see its use in high speed interconnections on existing and new fibers throught the backbone of the internet. The “see” being emphasised because most of us will not ever actually see it as they will be restricted to the higher layers of networking that most people will never see or experience directly. I guess what we will see hopefully in the longer term is reduced bandwidth costs and eventually that will allow cost effectively to deliver higher “usable” bandwidth to the home/business.

It is interesting that the speed of 40Gbps is not new but just new on the cheaper silicon substrate without needing the more costly and rare exotic materials they have previously used to achieve these high speeds.