Table 1 Summary of AI applications and models demonstrated on NeuRRAM
From: A compute-in-memory chip based on resistive random-access memory
Application | Dataset | Model architecture | Dataflow type | Activation precision | Number of parameters | Number of RRAMs used | Number of cores used | Average core utilization (%) |
|---|---|---|---|---|---|---|---|---|
Image classification | CIFAR-10 | ResNet-20 (CNN) | Forward | 3-bit unsigned, input image 4-bit unsigned | 274,461 | 553,524 | 48 | 17.6 |
MNIST | 7-layer CNN | Forwards | 3-bit unsigned | 23,170 | 46,664 | 16 | 4.5 | |
Voice recognition | Google voice command | 4 parallel LSTM cells | Recurrent + forwards | 4-bit signed | 281,392 | 570,048 | 36 | 24.2 |
Image recovery | MNIST | RBM | Forwards + backwards | Visible: 3-bit unsigned. Hidden: binary | 96,194 | 200,880 | 8 | 38.3 |
