Characterization
Profile the memory demands of neural network layers and accelerator types. Measure how a sensing application responds as other workloads execute.
Memory systems and security
Shared memory related security challenges in mobile computing platforms.
Processors on a mobile system on chip often share the same DRAM. When applications compete for memory bandwidth, the resulting slowdown can reveal information about other workloads.
The project studies how applications observe memory demand across accelerators, what those observations reveal, and how a system can reduce leakage.
Profile the memory demands of neural network layers and accelerator types. Measure how a sensing application responds as other workloads execute.
Examine whether contention signatures reveal network properties or enable communication between otherwise isolated workloads. MC3 studies a covert channel based on shared DRAM contention.
Investigate memory controller policies and randomization methods that obscure contention signatures, while considering the performance cost.

Memory contention also affects scheduling and performance prediction.