Automatic Allocation of Heterogeneous Resources
Ongoing research on assigning heterogeneous hardware resources from model definitions to improve neural network inference under tight resource limits.
This ongoing project develops methods for allocating heterogeneous hardware resources automatically from a model definition. The aim is to improve inference performance for larger neural networks while working within limited hardware budgets.
The research focuses on how different classes of hardware resources can be combined effectively rather than treating the target platform as uniform. By exploiting heterogeneous resources more systematically, the project aims to improve achievable performance without requiring manual allocation decisions at every stage of the design flow.
This work is still evolving, so the portfolio entry intentionally stays high level for now.