Fabless vs IDM for Data Centers: Silicon Photonics as a Disruptive Force?

I recently received a copy of a book entitled Silicon Photonics III (Amazon) and while perusing the book I was captured by the first chapter entitled ‘Silicon Optical Interposers for High-Density Optical Interconnects’. The chapter covered the work of a team in Japan on an idea they termed “on-chip servers” and “on-board data centers”. It brought back to mind an article on SemiWiki about ARM and TSMC and how the server market will be the next big battle between Fabless and IDMS (Data Center Fabless vs IDM). Could silicon photonics be the disruptive force that enables such a battle in the data centers?

The team that authored the referenced chapter are part of a collaborative project between the University of Tokyo, Photonics Electronics Technology Research Association (PETRA) and Advanced Industrial Science and Technology (AIST). Their idea is to combine the die that would comprise a server board into a single packaged element using a Silicon Optical Interposer. This is 2.5D / 3D stacking with a twist as it makes use of silicon photonics for the horizontal connections between the die. Stacked die are still vertically connected using through-silicon-vias (TSVs). Laser-diodes (LD), optical modulators (OM), photo-detectors (PD) and optical wave guides (OW) are integrated on the silicon interposer and are used by the digital ICs to communicate to each other. The group has demonstrated working prototypes with FPGAs flip-mounted on top of such an interposer running error-free inter-chip communications with bandwidth density of 30 Tbps/cm[SUP]2[/SUP] with a channel line rate of 20Gbps.

Conclusions from the authors were as follows. “Since the maximum lithography field size (stepper shot size) in area and signal I/O pad percentage out of total pads for CPUs have currently been 8.58 cm2 and 33% respectively and will be so in the future, we can obtain an overall inter-chip bandwidth at the level of several tens of Tbps by using silicon optical interposers, which is sufficient for the required bandwidth in the late 2010s or early 2020s.”

Take this one step further. The goal of the group in Japan is to by 2022 be able to show what they termed an “on-board Data Centers” which will combine multiple “on-chip servers” onto a single optical board. Communications between the on-chip servers both on a single board and between boards in a rack would be done through photonic optical communications. To enable this the group is working on what they call “optical I/O cores”. The first incarnation of these optical I/O cores will be designed to be mounted in an Active Optical Cable (AOC) module and then the next generation will be designed to be integrated around a host-LSI in the LSI’s package. The group has already demonstrated error-free data links using the Optical I/O Cores at 25 Gbps over a 300 meter multi-mode fiber (MMF). This implies that 100Gbps data links (25 Gbps x 4 channels) are feasible with a 3X longer reach than conventional solutions with less power consumption than conventional SR10, SR4, LR4 or PSM4 implementations even when you count the power required for the laser diodes to drive the optics. Will the IDMs such as Intel or Fujitsu use this type of architecture to keep the fabless guys out of the data centers? Or could a fabless company (say a Qualcomm with ARM cores) use this type of architecture as a wedge against the incumbent IDMs in the data centers? Intel is already making news with silicon photonics (Intel Announcements).

With all of that said, if we are truly going to see a battle between IDMs and Fabless in the server space we are going to need to start seeing some action out of the fabless foundries towards aggressively supporting production volume silicon photonics solutions.