Functional unit sharing between stacked processors in 3D integrated systems

The emerging Through Silicon Via (TSV) based 3D integration technology provides the means to stack two or more dies, enabling a low-latency interface between them. Apart of the immediate advantages of such an approach, e.g., short wires, it also opens research avenues for 3D organizations of computation platforms. In this line of reasoning we propose in this paper to share resources between stacked processors while focusing on Functional Units (FUs) inter-die sharing. The purpose of FU sharing is two fold: (i) it enables inexpensive fault tolerance by allowing, when possible, redundant instruction execution on idle FUs of processors located on other stack dies; and (ii) it can result in performance improvements by remotely executing instructions on idle FUs located on other dies in the 3D stack, when more instructions than locally available FUs are issuable. We evaluated the potential implications of our proposal on a 3D system built with two stacked processors (dies). In this case only a limited error detection capability is enabled and the experimental results indicate that our scheme covers on average 46% of the executed instructions. When performance improvement is targeted an average speedup of 6.9% can be achieved for the applications running on the considered two die stack.