Perfromance of FINN generated design

[pkdeep]

I have a few specific questions regarding the Latency performance of the design generated by FINN:

1. Does pruning impact Latency performance or resource utilization?
2. Does keeping "Bias" in float format impact performance?
3. How is the AXI Stream width limitation of 8192 calculated?:
   3.1 Like if I have 512x256 in my first layer, the input dimension is 512, and is quantized to 4 bits, and weight and activation are both 2-bits. How will the AXI Stream connection width be calculated: both in the case of decoupled memory and const mode?

Regards
Pradeep

[fpjentzsch]

Hi,

1. Only resource utilization (mostly for high parallelism). If you prune by actually reducing the number of channels in the model, or increase SIMD or PE because the reduced resource utilization gives you more room to parallelize, then it could also impact latency/throughput.
2. No
3. See the following functions to compute the stream widths:

   finn/src/finn/custom_op/fpgadataflow/matrixvectoractivation.py

   Line 262 in 39fb885

   def get_instream_width(self, ind=0):