Lighting Up The Cloud

In our rush to imagine a world populated with IoT devices, tech advances at the top end of this ecosystem (the cloud) don’t seem to get much airtime. But this isn’t because they are limited to modest refinements. As one example, there is active technology development in connectivity around fiber-based communications within the datacenter.

I always like to start with why changes are happing; an Intel shareholder update explains this quite well. These clouds (datacenters) are seeing three big shifts: massive growth in scale driving massively increased bandwidth demands, the same growth in in the scale of the building(s) containing the datacenter, requiring cabling reach up to 2km, and a change in traffic patterns from hierarchical enterprise flows (masters feeding/controlling slaves) to cross-system flows, emerging thanks to network function virtualization (NFV) and software-defined networking (SDN), requiring more flexible connectivity (less hierarchy).

Copper can’t keep pace with these demands, hence for some time fiber has been the preferred medium for connectivity, for all the usual reasons – low attenuation, security, material cost and more. The preferred light source is vertical-cavity surface-emitting lasers (VCSEL) operating at 850nm. But high-speed VCSELs typically have large spectral width, leading to dispersion in the fiber which limits reach to ~100m; they are also tend to higher power requirements than demanded by large datacenters.

Recent articles have suggested that silicon photonics may provide an answer. VCSELs are built in specialized devices distinct from the compute engine, but there is an intuitive appeal in building the optical link into the same chip (or package) as that engine, potentially improving performance, power and cost, also leveraging the technology and capabilities already developed for semiconductor manufacture. Intel and IBM certainly believe this. The Intel shareholder update I mentioned earlier shows silicon photonics can already meet the 2km reach requirement and is therefore able to bridge the gap between VCSEL reach and full datacenter size (I assume this would be as a result of narrower spectral width from the lasing source though Intel does not elaborate on this point). Better yet, the promise is that silicon photonics can drive 4 optical channels into one fiber, increasing capacity at no additional cost over cheaper single-mode fiber, where high performance VCSEL connections require multi-mode fiber.

So much for the promise – reality seems to be less clear. First, the silicon photonics domain appears to be as susceptible to hype as other domains; unsurprising since some companies have been investing in this area for years without serious (commercial) progress. On a brighter note, you can now buy devices (from Cisco, Mellanox and Infinera, for example) which use the technology, also the compelling needs remain, so this is hardly a failure. But silicon-based lasers still struggle with reliability (unlike VCSELs), which may be why Intel had a false-start release early in 2015, though they are now shipping production silicon. And the VCSEL folks are not standing still; research work has demonstrated the feasibility of VCSELs operating in the same range claimed for silicon photonics.

The net seems to be that VCSELs are proven today in production usage but have reach and power limitations which in principle could be overcome though not yet demonstrated outside the lab (?), whereas silicon photonics has promise but still seems to be teetering on the edge between promise and delivery. At the end of the day, this race may be decided as much by bill-of-materials and manufacturing cost as by technology, and that’s where silicon photonics should have the edge.

An article introducing the promise of silicon photonics can be found HERE. You can read the Intel shareholder update HERE. An article reviewing whether we are past the silicon photonics hype phase is HERE. A nice comparison of VCSELs versus silicon photonics can be found HERE. For an example of research in extending useful operating parameters of VCSELs, see HERE.

More articles by Bernard…