Reference Number: DEVCOM-134

Project Description

Students will work on adding additional functionality to the Neural Place and Route (NPR) software to improve performance. NPR is used to map an ONNX model of a deep neural network into a Field Programmable Neural Array (FPNA) chip architecture. Students will first learn the FPNA chip architecture in detail, then learn the inner workings of the NPR scripts, before being able to update NPR functionality. Updates will include 1) support of more types of ONNX neural network models, 2) more clear descriptions of constants, 3) Improve FPNA utilization (i.e. pack more compute into each tile), 4) better support for current networks (i.e. properly incorporate padding in convolutions), among other updates.

Technical Skills

• Debugging code
• Digital Circuit Design (Verilog or other HDL language)
• Extensive python coding experience