By Michael Klein, The University of Texas at Austin, Electrochemical Energy Laboratory and MRS Student Chapter
Amid the ubiquitous search for energy savings across all fields of modern society, reducing the power cost of computation looks to play an outsize role in limiting the expansion of energy demands into the foreseeable future. In the August issue of the MRS Bulletin, Lionel Kimerling, Dim-Lee Kwong, and Kazumi Wada look at the prospect of integrating microscale photonics into computer chips. As the authors point out, limitations imposed by transistors’ power draw has already shifted the direction of processor development from increasing clock speeds to increasing parallelism (the number of processing cores per chip).
This shift imposes two distinct problems for traditional chip architectures: the cores must be interconnected in an efficient, low-latency manner, and software must explicitly code for the appropriate inter-processor connections to take advantage of the increased parallelism. Both problems become exponentially worse with increasing numbers of processors. The authors propose that the use of integrated photonics can solve both issues. By utilizing a technique referred to as all-to-all computing (ATAC), each processor could communicate to every other processor on its own unique wavelength of light through multichannel photonic waveguides, rather than making individual connections from one processor to another. This architecture would decrease the energy demand of parallel processing by drastically reducing the consumption associated with core-to-core networking, possibly by a factor of several hundred. Concurrently, ATAC schemes dramatically simplify effective software utilization of parallel processing.
The article goes on to discuss the numerous challenges facing the realization of this vision. While all of necessary microscale optical devices have been developed in some form, the prospect of integrating high-volume manufacture with traditional microelectronic circuits at the requisite tolerances is daunting. For instance, in order for the ATAC scheme to work, waveguides with single-nanometer dimensional tolerances would need to be fabricated at a chip-scale. Promisingly, the authors discuss the success of the transformation of a 180-nm CMOS manufacturing line to fabricate microphotonics by the Institute of Microelectronics and GlobalFoundries in Singapore.
[Figure: Schematic representation of the complexity of scaling traditional inter-processor networking with a photonic ATAC approach. Courtesy of Cambridge University Press: [MRS Bulletin] L. Kimerling, et al., MRS Bull. 39 (8), 687-695, copyright 2014]