FLIK - Openvino on FPGA

[TNFPGA]

Hi,

I'm seeking your expertise regarding a few challenges I'm facing with the FLIK (FPGA Client Innovation Kit) on Windows.

My current workflow is as follows:

PyTorch → .ONNX → OpenVINO Model Optimizer → Inference on FPGA

However, I’ve encountered some limitations where certain operations are not supported on the FPGA, forcing me to use HETERO:FPGA,CPU. I’d like to confirm whether this behavior is expected or if there are configuration changes I should consider.

Specifically, the following operations fall back to CPU:

1. Slice operation (e.g., array[:, :, 10:20, :]) - on CPU

2. Element-wise multiplication with broadcasting, such as [1, 1, 32, 32] * [1, 48, 32, 32] - on CPU

3. Even with [1, 48, 32, 32] x [1, 48, 32, 32], - on CPU

Are these operations supported on FPGA in newer versions of OpenVINO? If I want it (the multiplication) to be executed on FPGA, what is the option on openvino?

Additionally, I'm currently using OpenVINO 2019R1 due to FLIK's installation requirements, which include the OpenVINO BSP and Toolkit 2019R1 for Windows with FPGA support. Since all related packages are tied to this specific version, I'm concerned about compatibility with newer versions.

Is there a path to upgrade to a newer OpenVINO version while maintaining support for the FLIK driver? As far as I understand, the driver may be tightly coupled with the BSP provided for 2019R1. Also, does the newer version of OpenVINO support INT8 precision for FPGA inference?

[Packages installed]

Then later on is the openvino toolkit

Have contacted Terasic and they direct me to here. Appreciate for any helps.  

Thank you in advance for your guidance.

[JohnT_Intel]

Hi,

This system is provided by Terasic. Please contact Terasic (sales@terasic.com) for FLIK support request.

Thanks.