BRAM usage per PE

[Joopi]

Hi,

In the FINN-R paper, the formula for calculating number of BRAMs required by the weight memory of a given layer is:

It explains that "This memory volume is split into separate memories, one for each processing element.".

However, in the "Evolutionary Bin Packing for Memory-Efficient Dataflow
Inference Acceleration on FPGA" paper, this is explained:

If PE * SIMD * W is the memory word width, then shouldn't the number of BRAMs required to implement this just be ceil((PE * SIMD * W) / BRAM_width) if we only consider the width dimension for simplicity?
Or does the Bin Packing paper just do a simplification without getting into FINN-specific implementation details that each PE requires a separate memory?

I am confused, could someone kindly explain the exact behavior?

[Joopi]

To elaborate a bit, my understanding from the FINN-R paper is that PE always multiplies the number of BRAMs, whereas from the Bin-Packing paper, PE and SIMD both scale the number of BRAMs equally. But maybe I'm just extrapolating too much?