As the per-transistor speed and energy efficiency improvements diminish, exploring alternative computing technologies and unconventional techniques is becoming critical to improving the performance and energy efficiency of general-purpose processors. Using analog circuits for computation is one such alternative technique aiming to deliver high performance within a very low power envelope. However, two of the main challenges that hinder the wide use of analog circuits for general-purpose computing: (1) the inherent computational inaccuracy in these models and (2) the lack of easy-to-use programming models. On the other hand, conventional von Neumann computing carries a long tradition of programming abstractions that has enabled the recent exponential advances in the computing industry. However, these traditional architectures do not provide the same degree of efficiency as the analog systems. Due to the inherent efficiency of analog circuits, they can provide significant leaps in efficiency of computer systems when applications can tolerate inexact and approximate computation. There are indeed many emerging applications, including machine learning, big data analytics, augmented reality, speech recognition, signal processing, robotics, health monitoring, wearable devices, IoT and many more that can potentially benefit from analog computation. These approximable applications can tolerate inexact computation in substantial portions of their execution. Neural transformation [1,2], for example, is one of the possible ways to utilize the inherent efficiency of the analog computation.
In summary, I think analog computation has a lot of unexplored potential and is one of the promising ways to provide not only higher efficiency but more importantly to unlock new capabilities in the computing landscape. The community needs more research across the full system stack to reap the full potential of analog computation. The good news is that computer architecture community [2,3] and many funding institutes such as DARPA (UPSIDE Project)  have started to explore the analog computation potentials. I believe that analog computation is a golden opportunity for all of us and we need to explore groundbreaking approaches in this domain which for sure lead to significant improvements in the efficiency of the computation. I am looking forward to see novel algorithms, techniques and designs in the domain of analog computation.
 Hadi Esmaeilzadeh et al. "Neural Acceleration for General-Purpose Approximate Programs" - MICRO 2012.
 Renée St. Amant, Amir Yazdanbakhsh et al. "Toward General-Purpose Code Acceleration with Analog Computation" - ISCA 2014.
 Amir Yazdanbakhsh et al. "Neural Acceleration for GPU Throughput Processors" - MICRO 2015.
 Unconventional Processing of Signals for Intelligent Data Exploitation (UPSIDE).
comments powered by Disqus