Resistive RAM (RRAM) is a promising option in the light of diminishing returns from CMOS technology scaling, however, RRAM devices are inherently unstable and require amortization effort to compensate for process variations, wear-off, and stochastic effects during read and write. Khunjerab gears to exploit resistive memory for ML applications, in particular with tradeoffs regarding performance, power consumption, and reliability.
Thus, in the Khunjerab project we are most interested in codesigning applications and RRAM while not gearing towards upmost reliability, instead we gear to identitfy the most optimal combination within the tuple of performance, power, and reliability (PPR). For instance, some applications might tolerate certain inaccuracies during read and write, while others require that memory operations are certain.
It is a joint project with colleagues from Xi’an University of Technology (XAUT), China, the Instiute for Micro-Electronics (IMECAS) at Chinese Academy of Sciences, and the Compute Systems Group (CSG) of Heidelberg University. While XAUT and IMECAS will contribute their experience and expertise in designing innovative RRAM and related devices, CSG will contribute expertise in computer architecture, machine learning, compilers and related compute stack components, and hardware design.
Khunjerab has receided funding from the Deutsche Forschungsgemeinschaft (DFG, German national funding, Projektnummer 449797478) and National Natural Science Foundation of China (NSFC), covering a period of three years. The project has started in May 2021.
Holger Fröning, Computer Systems Group (CSG), Institute of Computer Engineering, Heidelberg University, Germany (Co-PI)
Jiawei Zhang, Xi’an University of Technology (XAUT), Xi’an, China (Co-PI)
Kai Xi, Instiute for Micro-Electronics (IMECAS), Chinese Academy of Sciences, Beijing, China (Co-PI)
Yannick Emonds, Computer Systems Group (CSG), Institute of Computer Engineering, Heidelberg University, Germany (PhD student)
Holger Fröning: holger.froening (at) ziti.uni-heidelberg.de