neurosim/DNN_NeuroSim_V1.4

SRAM onoffratio setting suggestion

Opened this issue · 3 comments

As the manual says, the SRAM onoffratio might have issues. Do you have any suggestion for the setting?

We do not recommend using NeuroSim as-is to determine the inference accuracy using analog SRAM compute-in-memory. We currently only support the inference accuracy estimations for memories that follow Ohm's law and Kirchoff's current law to compute the MAC operation (e.g. RRAM, FeFET). Analog SRAM CIM uses a different mechanism for the MAC operation and thus are not currently supported in our inference estimation framework. This is why in the manual we set the on/off ratio to infinity for the SRAM benchmarks.

So, what is the setting to re-generate the table in the manual (SRAM)? Or we just use the SRAM to compare with NVM CIM?

You can use --onoff float('inf') to generate the results from the table. Using a sufficiently high number (e.g. 1M) should also have the same result.