Inquiry Regarding FFT-Based Neural Network Support on Hailo

General Discussion
1

Hello Team,
I am currently investigating FFT-based neural network approaches, particularly block-circulant matrix methods (e.g., CirCNN), and evaluating their suitability for deployment on modern AI accelerator hardware.
I would appreciate your insights regarding your platform’s support for such workloads.

Specifically, I would like to ask:

  1. Does your accelerator architecture provide any dedicated FFT hardware support, specialized FFT instructions, or FFT-oriented processing pipelines?
  2. Are there optimized FFT kernels, libraries, or software components available within your SDK or software stack?
  3. Does the platform support custom operators or low-level kernel development that would allow implementation of FFT and IFFT operations?
  4. Is native complex-number support available, or would real and imaginary components need to be handled separately?
  5. In your opinion, would FFT-based neural network models, such as block-circulant networks, be an efficient use case for your hardware compared to conventional matrix-multiplication-based models?

I am particularly interested in understanding both the architectural capabilities and software support available for implementing and evaluating FFT-based neural network layers.

Thank you very much for your time and assistance.

Best regards,
Sowmiya Arumugam

2

Welcome to the Hailo Community!

To run a model (or part of a model) on the Hailo architecture, it must be converted into a HEF file using the Hailo Dataflow Compiler. The Hailo Dataflow Compiler User Guide describes the supported layers and parameters. You can download it from the Developer Zone or find it in the Hailo AI Software Suite Docker image under:

/shared_with_docker/doc/

We have added support for new layers over time to accommodate newer models. However, support for features that require significant development effort is evaluated on a case-by-case basis and typically requires a clear business justification.

If you have a product opportunity that may justify such an investment, please feel free to send me a PM.