Advanced Thermal Management using Approximate Computing and On-Chip Thermoelectric Cooling

In this work, we investigate the effectiveness of approximate computing and on-chip thermoelectric cooling on mitigating and managing the on-chip temperatures. On the first hand, approximate computing has emerged in the last decade as an attractive computing paradigm that offers a powerful trade-off between power and accuracy. For deep learning ap-plications, approximations in computing may not always lead to an observable loss in accuracy. This largely depends on the sensitivity of the executed DNN models. On the other hand, on-chip cooling using novel ultra thin-film Thermoelectric (TE) devices has also emerged as an attractive powerful means for heat dissipation to suppress excessive on-chip power densities. Our investigations are done using commercial ANSYS tool flows that employ accurate Finite Elements Analysis (FEA). This enables us to accurately study the Peltier's effect (for cooling purposes) as well as the Seebeck's effect (for energy harvesting purposes) demonstrating the promise and effectiveness of on-chip cooling using Thermoelectric devices.