Modular, co-design ML accelerator for FPGAs. The accelerator will explore characeristics of the model and the co-design will give feedback to model optimizations (quantization, pruning, etc.). The result includes different modules to efficiently implement different layers of a CNN. In addition the development framework includes tools for performance estimation and design space exploration. The first target applications include CNN inference (MobileNET).
Addressing target audiences and expressing needs
- Collaboration
- Fellowship to advance my/our research
Collaboration with other Universities and research institutesand/or fellowships to continue the work in the tool and then further support to bring the tool to a level that it can be made avialable to third parties.
- Research and Technology Organisations
- Academia/ Universities
R&D, Technology and Innovation aspects
Initial stage of tool development. Initial results published in literature.
Next steps include the broading of the scope by applying the same concepts to a broader set of ML models. Additional steps include making the tool available to third parties.