Previously, the sheer number of different device simulations required to properly analyze the statistics of device variability have made solutions impractical with conventional computing resources. However, the availability of HPC resources such as ScotGrid have made such simulations viable. Computation on this scale, however, is not without problems, and whilst using GANGA as a submission interface alleviated difficulties associated with large scale job submission to Globus other difficulties arose resulting in a considerable proportion of rogue jobs. This complicated data management, as it is important to avoid the generation of duplicate devices in order to preserve correct ensemble statistics. In consultation with ScotGrid admins, we have managed to overcome some of these difficulties to produce the first device ensemble of this scale. Work is also on-going exploiting OMII-UK middleware and in particular the use of technologies such as GridSAM for job submission and management.
Provide a set of generic keywords that define your contribution (e.g. Data Management, Workflows, High Energy Physics)
NanoCMOS transistors, device variability, numerical simulation, OMII-UK, GANGA
URL for further information:
1. Short overview
Progressive scaling of CMOS devices has driven the phenomenal success of the semiconductor industry. Silicon technology has now entered the nanoCMOS era with 40nm gate length transistors in production. However the semiconductor industry faces many fundamental challenges which will affect the design of future integrated circuits. The NanoCMOS project aims to apply eScience technologies to this problem. We describe our experiences (good and bad) with EGEE and OMII-UK technologies for this purpose.
4. Conclusions / Future plans
Simulation of a 100,000 device ensemble has consumed a considerable amount of computing power - over 11 years of CPU time over the course of approximately 6 weeks. In order to fully understand device variability we must now consider additional physical effects, and simulation of smaller devices. Currently we plan to proceed with the generation of large statistical ensembles for 25, 18, 13 and 9nm devices in order to examine variability at extreme levels of scaling.