Data centers are morphing into computing singularities, albeit large ones. Silicon photonics will hasten that process. The reason why begins with Moore's Law.
Gordon Moore's prediction known as Moore's Law -- "The number of transistors incorporated in a chip will approximately double every 24 months." -- has been uncanny in its accuracy since he made it in April 1965. That didn't stop pundits from saying Moore's Law had a nice run, but like all good things, it was coming to an end. The pundits' prediction was erroneous, thanks to Intel (the company Moore co-founded). The reason is light, or more accurately photons.
The problem photons overcome
Moore's Law requires scientists and engineers to continually figure out how to pack larger quantities of transistors and support circuitry into chips. It's a challenge, but not as difficult as figuring out what to do about the by-products of shoving electricity through an ever-more dense population of chips: heat buildup, current leakage, and crosstalk between adjacent wire traces.
Multi-core technology breathed new life into Moore's Law, but only for a short time. Using copper wires to transmit the digital information becomes the limiting factor. This MIT Technology Review 2005 article explains why copper wires were no longer good enough. "The problem is that electrical pulses traveling through a copper wire encounter electrical resistance, which degrades the information they carry," states author Robert Service. "As a result, data bits traveling through copper must be spaced far enough apart and move slowly enough that devices on the other end of the wire can pick them up."
That challenge becomes evident when walking through a data center, because most, if not all, copper-based Ethernet runs have been replaced with fiber optics. Using existing fiber-optic technology will not help Moore's Law -- that requires a new technology called the silicon laser.
Fast forward to 2009
Intel's Photonics Technology Laboratory in 2009 mastered the silicon laser. "We have done all the things that skeptics said we could not," mentions Intel Fellow Mario Paniccia in this SPIE article. "We have got beyond the proof-of-principle stage. Now we're putting it all together so that Moore's Law can extend for decades into the future."
The article goes on to explain how Paniccia and his team created high-speed silicon modulators and photodetectors so small they will fit on chips. The slide below depicts the two devices and their interconnections.
Innovations since 2009
Since 2009, Intel introduced:
- 50 Gigabit per second silicon-based optical data connection. The world's first silicon-based photonics link running at 50 Gbps, using technology that combines fiber-optic attributes with silicon manufacturing processes.
- Photonics technology operating at 100 gigabits per second. This is an integrated module including silicon modulators, detectors, waveguides, and circuitry.
- Optical PCI Express server. Fujitsu and Intel showcased new silicon-photonic connections that allow PCI cards to be moved off the main board, which creates shared pools of compute and storage, enhances cooling flexibility, and lowers costs by moving hot components farther apart.
Moving data centers to a single computing entity
One by-product of securing Moore's Law for the foreseeable future will be the complete redesign of data centers. Racks and racks of heat-spewing servers will be replaced by efficient, discrete components that are connected using silicon photonics.
For example, in 2013, Intel and Facebook released information about using silicon photonics at the rack level. "Intel and Facebook are collaborating on a new disaggregated, rack-scale server architecture that enables independent upgrading of compute, network, and storage subsystems that will define the future of mega-datacenter designs for the next decade," said Justin Rattner, Intel's then CTO. "The disaggregated rack architecture includes Intel's new photonic architecture...that enables fewer cables, increased bandwidth, farther reach and extreme power efficiency compared to today's copper based interconnects."
Disaggregated refers to separating compute, storage, networking, and power distribution resources into modules housed in the rack. "Traditionally, a server within a rack would each have its own group of resources," according to the press release. "When disaggregated, resource types can be grouped together and distributed throughout the rack, improving upgradability, flexibility and reliability while lowering costs."
So look for a two-pronged attack on copper in the data center. First, what Intel considers "pluggable" -- its MXC connector and new technology will revamp connections even as short as five inches. Second, embedded technology using silicon photonics will supply high-speed optical links to and from the processor.
I have written that data-center technologists are striving to morph data centers into a virtual and physical singularity. It appears that silicon photonics will help them reach their goal.