Briefing: Silicon photonics
Part 1: Optical interconnect
Luxtera demonstrated a 100 Gigabit QSFP optical module at the OFC/NFOEC 2013 exhibition.
"We're in discussions with a lot of memory vendors, switch vendors and different ASIC providers"
Chris Bergey, Luxtera
The silicon photonics-based QSFP pluggable transceiver was part of the Optical Internetworking Forum's (OIF) multi-vendor demonstration of the 4x25 Gigabit chip-to-module interface, defined by the CEI-28G-VSR Implementation Agreement.
The OIF demonstration involved several optical module and chip companies and included CFP2 modules running the 100GBASE-LR4 10km standard alongside Luxtera's 4x28 Gigabit-per-second (Gbps) silicon photonics-based QSFP28.
Kotura also previewed a 100Gbps QSFP at OFC/NFOEC but its silicon photonics design uses two chips and wavelength-division multiplexing (WDM).
The Luxtera QSFP28 is being aimed at data centre applications and has a 500m reach although Luxtera says up to 2km is possible. The QSFP28 is sampling to initial customers and will be in production next year.
100 Gigabit modules
Current 100GBASE-LR4 client-side interfaces are available in the CFP form factor. OFC/NFOEC 2013 saw the announcement of two smaller pluggable form factors at 100Gbps: the CFP2, the next pluggable on the CFP MSA roadmap, and Cisco Systems' in-house CPAK.
Now silicon photonics player Luxtera is coming to market with a QSFP-based 100 Gigabit interface, more compact than the CFP2 and CPAK.
The QSFP is already available as a 40Gbps interface. The 40Gbps QSFP also supports four independent 10Gbps interfaces. The QSFP form factor, along with the SFP+, are widely used on the front panels of data centre switches.
"The QSFP is an inside-the-data-centre connector while the CFP/CFP2 is an edge of the data centre, and for telecom, an edge router connector," says Chris Bergey, vice president of marketing at Luxtera. "These are different markets in terms of their power consumption and cost."
Bergey says the big 'Web 2.0' data centre operators like the reach and density offered by the 100Gbps QSFP as their data centres are physically large and use flatter, less tiered switch architectures.
"If you are a big systems company and you are betting on your flagship chip, you better have multiple sources"
The content service providers also buy transceivers in large volumes and like that the Luxtera QSFP works over single-mode fibre which is cheaper than multi-mode fibre. "All these factors lead to where we think silicon photonics plays in a big way," says Bergey.
The 100Gbps QSFP must deliver a lower cost-per-bit compared to the 40Gbps QSFP if it is to be adopted widely. Luxtera estimates that the QSFP28 will cost less than US $1,000 and could be as low as $250.
Optical interconnect
Luxtera says its focus is on low-cost, high-density interconnect rather than optical transceivers. "We want to be a chip company," says Bergey.
The company defines optical interconnect as covering active optical cable and transceivers, optical engines used as board-mounted optics placed next to chips, and ASICs with optical SerDes (serialiser/ deserialisers) rather than copper ones.
Optical interconnect, it argues, will have a three-stage evolution: starting with face-plate transceivers, moving to mid-board optics and then ASICS with optical interfaces. Such optical interconnect developments promise lower cost high-speed designs and new ways to architect systems.
Currently optics are largely confined to transceivers on a system׳s front panel. The exceptions are high-end supercomputer systems and emerging novel designs such as Compass-EOS's IP core router.
"The problem with the front panel is the density you can achieve is somewhat limited," says Bergey. Leading switch IC suppliers using a 40nm CMOS process are capable of a Terabit of switching. "That matches really well if you put a ton of QSFPs on the front panel," says Bergey.
But once switch IC vendors use the next CMOS process node, the switching capacity will rise to several Terabits. This becomes far more challenging to meet using front panel optics and will be more costly compared to putting board-mounted optics alongside the chip.
"When we build [silicon photonics] chips, we can package them in QSFPs for the front panel, or we can package them for mid-board optics," says Bergey.
"If it [silicon photonics] is viewed as exotic, it is never going to hit the volumes we aspire to."
The use of mid-board optics by system vendors is the second stage in the evolution of optical interconnect. "It [mid-board optics] is an intermediate step between how you move from copper I/O [input/output] to optical I/O," says Bergey.
The use of mid-board optics requires less power, especially when using 25Gbps signals, says Bergey: “You dont need as many [signal] retimers.” It also saves power consumed by the SerDes - from 2W for each SerDes to 1W, since the mid-board optics are closer and signals need not be driven all the way to the front panel. "You are saving 2W per 100 Gig and if you are doing several Terabits, that adds up," says Bergey.
The end game is optical I/O. This will be required wherever there are dense I/O requirements and where a lot of traffic is aggregated.
Luxtera, as a silicon photonics player, is pursuing an approach to integrate optics with VLSI devices. "We're in discussions with a lot of memory vendors, switch vendors and different ASIC providers," says Bergey.
Silicon photonics fab
Last year STMicroelectronics (ST) and Luxtera announced they would create a 300mm wafer silicon photonics process at ST's facility in Crolles, France.
Luxtera expects that line to be qualified, ramped and in production in 2014. Before then, devices need to be built, qualified and tested for their reliability.
"If you are a big systems company and you are betting on your flagship chip, you better have multiple sources," says Bergey. "That is what we are doing with ST: it drastically expands the total available market of silicon photonics and it is something that ST and Luxtera can benefit from.”
Having multiple sources is important, says Bergey: "If it [silicon photonics] is viewed as exotic, it is never going to hit the volumes we aspire to."
Part 2: Bell Labs on silicon photonics click here
Part 3: Is silicon photonics an industry game-changer? click here