Ed's Threads

Monthly Archives: December 2015

CMOS-Photonic Integration Thermally Sensitive

As published in the journal Nature, CMOS transistors have been integrated with optical-resonator circuits using complex on-chip sensors and heaters to maintain temperature to within 1°C. While lacking the laser-source, these otherwise-fully-integrated solutions demonstrate both the capability as well as the limitation of trying to integrate electronics and photonics on a single-chip. The Figure shows a simplified schematic cross-section of the device.

Full chip cross-section (not to scale) from the silicon substrate to the C4 solder balls, showing the structures of electrical transistors, waveguides, and contacted optical devices. The minimum separation between transistors and waveguides is <1 μm, set only by the distance at which evanescent light from the waveguide begins to interact with the structures of the transistor.

Full chip cross-section (not to scale) from the silicon substrate to the C4 solder balls, showing the structures of electrical transistors, waveguides, and contacted optical devices. (Source: Nature)

Lead author Chen Sun—affiliated with UC Berkeley and MIT, as well as with commercial enterprise Ayar Labs, Inc.—developed the thermal tuning circuitry, designed the memory bank, implemented the ‘glue-logic’ between various electronic components, and performed top-level assembly of electronics and photonics. The main limitation is the temperature control, since deviation by more than 1°C results in loss of coupling that otherwise provides for P2M/M2P transceivers:

* Waveguide Loss – 4.3 dB/cm,
* Tx and Rx Data Rate – 2.5 Gb/s,
* Tx Power – 0.02 pJ/bit,
* Rx Power – 0.50 pJ/bit, and
* Ring Tuning Control Power – 0.19 pJ/bit, so
* Total power consumption = 0.71 pJ/bit.

The Register reports that this prototype has a bandwidth density of 300 Gb/s per square millimetre, and needs 1.3W to shift a Tb/s straight from the die to off-chip memory. A single chip integrates >70 million transistors and 850 photonic components to provide microprocessor logic, memory, and interconnect functions.

—E.K.

Apple Fab Speculation

Apple Corp. recent purchased an old 200mm-diameter silicon wafer fab in San Jose capable of creating as small as 90nm device features. Formerly owned and operated by Maxim, the US$18.2M purchase reportedly includes nearly 200 working fab tools. Some people outside the industry have speculated that Apple might use this fab to do R&D on the A10 or other advanced logic chips, but this old tool-set is completely incapable of working on <45nm device features so it’s useless for logic R&D.

As reported at EETimes, this old fab could be used for the R&D of “mixed-signal devices, MEMS and image sensors and for work on packaging.” Those who know do not speak, while those who speak do not know…I do not know so I’m free to join the public speculation. Mixed-signal and MEMS processing would require major re-tooling of the line, but this 15-20 year-old tool-set is nearly turn-key for wafer-level packaging (WLP). With minimal re-tooling, this line could produce through-silicon vias (TSV) or through-mold vias (TMV) as part of Fan-Out WLP (FO-WLP).

Our friends at ChipWorks have published a detailed tear-down analysis of the System-in-Package (SiP) used in the first generation Apple Watch; it contains 30 ICs and many discretes connected by a 4-layer printed circuit board (PCB). Significant power and performance improvements in mobile devices derive from stacking chips in such dense packages, and even greater improvements can found in replacing the PCB with a silicon interposer. With Apple pushing the limits on integrating new functionalities into all manner of mobile devices, it would be strategic to invest in WLP R&D in support of application-specific SiP design.

—E.K.