Table 1 Summary of various DNN-to-IMC mapping schemes and quantized weight implementation methods proposed in the literature.
From: Device quantization policy in variation-aware in-memory computing design
HP lab12 | ISAAC13 | ASU14 | XNOR-SRAM15 | XNOR-RRAM16 | IBM17 | PRIME18 | NTHU19 | |
|---|---|---|---|---|---|---|---|---|
Mapping | N-IMC | N-IMC | N-IMC | C-IMC | C-IMC | WS-IMC | WS-IMC | WS-IMC |
Weight | S-MLC | D-MLC | A-MLC | S-MLC | S-MLC | S-MLC | D-MLC | S-MLC |
Device | RRAM | N/A | RRAM | SRAM | RRAM | PCM | RRAM | RRAM |
Cell precision | 5-bit | 2-bit | 6-bit | 1-bit | 1-bit | 5-bit | 4-bit | 1-bit |
Weight precision | 5-bit | 16-bit | 6-bit | 1-bit | 1-bit | 6-bit | 8-bit | 1-bit |
Activation precision | 4-bit | 8-bit | N/A | 1-bit | 1-bit | 8-bit | 6-bit | 1-bit |
