Personal tools

Data Mapping for Unreliable Memories

From iis-projects

Revision as of 13:56, 9 February 2015 by Weberbe (talk | contribs)
(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)
Jump to: navigation, search
Top: Digital communication system employing BPSK transmission over an AWGN channel with unreliable memory. Bottom: Bit-error rate performance of the system assuming convolutional coding for different data representations.




Christoph Roth
Christian Benkeser
Georgios Karakonstantis (EPFL)
Andreas Burg (EPFL)
Christoph Studer (Rice University)


Telecommunications Circuits Laboratory of EPFL
Digital Signal Processing Laboratory of Rice University


The continuous shrinkage of semiconductor devices during recent years has led to the enormous success of digital signal processing (DSP) systems. Such an evolution was – up to now – relying on the assumption that the underlying hardware is able to perform computations and store data in a 100% reliable manner. However, it is now becoming apparent that such a trend may come to an end due to the increasing effect of semiconductor-process variability as well as reliability issues that threaten the correct circuit functionality, especially for CMOS technology nodes beyond 45 nm. Consequently, future DSP systems must provide robustness on algorithm and application level to the presence of reliability issues. In this project, we have investigated the impact of defects in memories, which are particularly susceptible to process variations, on the performance of DSP systems. We have found that this impact can be reduced by choosing data representations that are different from the ones typically employed in digital integrated circuits, such as 2’s complement or sign-magnitude number formats. Based on this observation, we have developed a novel framework to analyze the impact of data representations on the performance of unreliable systems. As a proof of concept, we have applied this framework to a practical coded communication system. We have observed that the deployment of optimized data representations (based on application-specific cost functions) substantially increases the robustness of the system to unreliable memory operation, compared to the data representations most commonly used.


  • C. Roth, C. Benkeser, C. Studer, G. Karakonstantis, A. Burg, "Data Mapping for Unreliable Memories", 50th Annual Allerton Conference on Communication, Control, and Computing, Monticello, Illinois, USA, 1-5 Oct 2012