Intel’s ecosystem validates silicon photonics interconnect

OVUM VIEW

Summary

Intel, along with an impressive list of collaborators, revealed a photonic architecture for server racks at the Open Compute Project Summit in Santa Clara, California, on January 17. A novel optical connector from Corning is included, though ancillary details on Intel’s silicon photonics are yet to come. Quanta will be building a disaggregated server for Facebook based on this architecture. This is the event we have been waiting for since Intel went into stealth mode with its silicon photonics technology two years ago. While it is only one piece of the Open Compute architecture, the implications for the photonics industry are significant. The creation of an ecosystem comprised of large incumbent suppliers denotes mainstreaming of optical interconnect and silicon photonics.

Silicon photonics enables future-proof single-mode parallel fiber

While pluggable front-panel optical transceivers are commonly used for data center links on the scale of tens or hundreds of meters, Intel’s new architecture moves optics into the system backplane. Embedded parallel optics based on multimode fiber and VCSEL photonics (e.g. SNAP12 microPOD, called “miniSNAP12” in Ovum’s Datacom Forecast) have been used in intra-system interconnect for years, but applications have remained limited and proprietary. As communication rates increase, the reach of media such as copper, active copper, and multimode shrink to shorter and shorter distances over time. Recently, with the rise of 40GbE SR4 in the data center, multimode ribbon fiber has become accepted.

Single-mode ribbon fiber has the potential to be a single high-density cabling infrastructure that works from intra-rack to inter-building in a mega data center complex, and that moreover is future-proof. The missing pieces have been economical parallel ribbon cables, particularly connectors, and low-cost, low-power single-mode array transmitters. Intel’s announcement addresses both. A new cost-reduced connector developed by Corning can handle up to 4 rows of 16 fibers in a similar footprint – and use case – to today’s MPO/MTP ribbon connector. Silicon photonics provides the single-mode active transmitter and receiver components.

New hope for optical interconnect via open supply chain

Hence this announcement, though it does not directly compete with front-panel transceivers, does strengthen the outlook for a common technology base that ultimately spans even further from on-chip to WAN distances. In our Silicon Photonics: A Tale of Two Markets report last year, we described a landscape in which the supply chain had gaps between optics vendors and system customers. For inter-system, conventional optics remains a more comfortable choice for customers than silicon photonics. At the short-distance extreme, computer architects believe photonic chip-scale interconnect will be needed, but no vendors – certainly none from the established computer supply chain – are yet offering it. The Intel ecosystem offering, in contrast, falls into the sweet spot between these two markets – intra-rack but not yet integrated into the chip package – and does bring together the suppliers and customers.

Application pull, not technology push

Intel is ramping commercial production of its silicon photonics, which will go into Facebook’s servers manufactured by Quanta. The Open Compute Project was spearheaded by Facebook to gather industry momentum around its modular, scalable architecture directions, explicitly asking for improved designs down to the component technology. What remains to be seen is how Intel will make its technology available to the broader market.

This announcement is particularly promising as it centers on the system benefits, not the technology. The disaggregated design takes advantage of the interconnect’s speed and low latency to separate the different types of resources in a modular design without sacrificing performance. Differently customized servers, such as storage-heavy ones for Big Data or compute-centric ones for high-performance computing, could be built from common elements. This separation also allows upgrade cycles for each type of resource – memory, processing, storage, networking – to be decoupled. Such flexibility suggests the architecture might even be extended beyond the usual domain of servers for even more specialized systems.

APPENDIX

Author

Karen Liu, Principal Analyst, Components

karen.liu@ovum.com

Further reading

Silicon Photonics: A Tale of Two Markets, TE003-000522 (September 2012)

Total OC Forecast Spreadsheet: 2012–17, TE003-000520 (August 2012)

Disclaimer

All Rights Reserved.

No part of this publication may be reproduced, stored in a retrieval system, or transmitted in any form by any means, electronic, mechanical, photocopying, recording, or otherwise, without the prior permission of the publisher, Ovum (an Informa business).

The facts of this report are believed to be correct at the time of publication but cannot be guaranteed. Please note that the findings, conclusions, and recommendations that Ovum delivers will be based on information gathered in good faith from both primary and secondary sources, whose accuracy we are not always in a position to guarantee. As such Ovum can accept no liability whatever for actions taken based on any information that may subsequently prove to be incorrect.