Report of resource consumption

[MFJI]

Hello all,

I found that the resource consumption from the FINN report is more than the resources that the board I set has. Does anyone know what is the reason for this? Thanks.

[MFJI]

I am using the board alveo U200, and I found that there is no board file of the alveo U200 in Vivado 2022.1. Is this the thing that causes the problem? Is the report in step "step_generate_estimate_reports" serious enough?

And also, is it possible for me to decide the type of RAM to use? For example, I would like to use BRAMs instead of URAMs?

[auphelia]

Hi @MFJI ,

The resources that FINN reports can be on three different levels:

• FINN resource estimations: Calculates resource estimations based on FINN-R: An End-to-End Deep-Learning Framework for Fast Exploration of Quantized Neural Networks
• HLS resources: Resource estimations from the Vitis HLS synthesis
• Post synthesis resources: Resources reported by Vivado synthesis

I assume you have used the resource estimation builder step, which generates resource estimations before any synthesis was run. As the name says, these are "estimations" and might over-estimate your actual resource usage.

You can specify which memory to use by specifying the ram_style node attribute. It can be specified using the folding_config.json. You can find an example here: https://github.com/Xilinx/finn/blob/main/src/finn/qnn-data/build_dataflow/folding_config.json

The options are:

• auto -> let Vivado/Vitis decide
• block -> BRAM
• distributed -> LUTRAM
• ultra -> URAM