Extended Data Table 1 Comparison of NVM-based processors and macros

From: A mixed-precision memristor and SRAM compute-in-memory AI processor

  1. *1Estimated by throughput/area.
  2. *2Based on operation ratio of Fig. 4a using the MobileNet-v2 model trained for ImageNet with balanced configuration.
  3. *3Based on operation ratio of Fig. 4a using the ResNet-20 model trained for CIFAR-10 with balanced configuration.
  4. *4Normalized by operation ratio of each mode and bit number of FP16 and INT8.
  5. *5The highest value is based on CIFAR-10, whereas the lowest value is based on ImageNet.
  6. *6One operation (OP) represents one multiplication or one addition. Data from refs. 1,2,3,9,10,11,15,23,24,27.