Intel Corp showed off its silicon photonics technology at this week’s Open Compute Summit (OCP) in Santa Clara, California. The event is a forum for industry players to share technology and ideas for the betterment of datacenter hardware. In this case, betterment means greater efficiency, less complexity and more open standards. Intel believes its silicon photonics technology can drive all of these goals, especially the first two.
The idea behind the silicon photonics is to bring optical communication into the realm of semiconductors, thus leveraging its natural advantages of density, power efficiency, and scalability. Integrating photonics on-chip will also allow for much higher data transfers between processors and devices than is possible with electronics-based optical componentry. All of that translates into hardware that needs fewer parts and is less expensive to build and run.
The OCP presentation, which was delivered by Intel CEO Justin Rattner, featured an Intel-designed prototype “photonic rack” built by Quanta Computer. According to the press release Intel has moved its silicon photonics work beyond the R&D stage, and the company has produced engineering samples that can run at the 100 Gbps speeds. The OCP set-up incorporated this 100G technology, which “enables fewer cables, increased bandwidth, farther reach and extreme power efficiency compared to today’s copper based interconnects.”
The prototype also makes use of Ethernet switch silicon, presumably based on technology from Fulcrum, which Intel acquired back in 2011. The motherboard design is such that the photonic rack will support Intel’s Xeon and the next-generation Atom processors (“Avoton”). It was unclear how the photonics capability was intended to be integrated, although Rattner’s presentation slides showed motherboard mock-ups for both Ethernet and PCIe connectivity.
The work is part of a collaboration between Intel and Facebook to build future server racks that enable the disaggregation of compute, switching and storage. The rationale is that once you have the optical communications on-chip, data transfers between computers or between computers and storage become much faster and simpler and require a good deal less energy. That means it makes more sense to build separate boxes for compute and storage, since transfer speed (and latency) is less of a limiting factor. At some point, Intel believes the technology can be used to decouple compute and memory as well.
Although this particular effort was directed at the needs of Facebook and similar types of ultra-scale datacenter applications, silicon photonics technology is also expected to figure prominently in the products aimed at the high performance computing market. And given Intel’s plans to also integrate network controllers across their processor line, it’s not too hard to see where the company is headed with all of this.