Please enable JavaScript.
Coggle requires JavaScript to display documents.
In-Memory Computing - Coggle Diagram
In-Memory Computing
Mapping weights to XBAR
Flatten filter weights
shafiee2016isaac
Compilation of models for CM
Data dependencies
Polyhedral compilation
kornilios2020compile_nn_for_cm
Patitioning and lowering
kornilios2020compile_nn_for_cm
Signed arithmetics support
XBAR inputs
2's complement
shafiee2016isaac
In XBAR
representation with bias
Architectures
Hardware architectures
ISAAC
shafiee2016isaac
IBM
kornilios2020compile_nn_for_cm
Analog/Digital Conversion Overhead
Encoding to reduce ADC resolution
shafiee2016isaac
Reduce use of ADC
nag2018newton
Reducing intra-IMA communication
Use Karatsuba's algorithm to reduce the complexity of matrix multiplication to reduce the use of the ADC.
Embed shift-and-add units in the HTree (network between cells)
Reducing inter-IMA communication
Use Strassen's algorithm (linear algebra manipulations) to use less IMAs thus reducing ADC usage
Tile architectures Optimizations
Heterogeneous tiles for convolution and classifiers
nag2018newton
Training techniques for CM
Generic Software training
Injecting noise
in Forward prop
joshi2020dnn_inference_pcm
Mixed-precision DL training
eleftheriou2019dl_acceleration_on _cm
nandakumar2020mixedprecision_dl_on_cm
Accuracy Retention Optimizations
Calibration techniques
joshi2020dnn_inference_pcm
during Batch Normalization
Global drift compensation (GDC): Scaling the crossbar output by a 1/a factor computed while device is in IDLE mode. Scaling can be merged to other computing stages like Batch Normalization.
Adaptive batch normalization statistics update (AdaBS): Recomputing and updating mean and variance for every layer with Batch Normalization. Mean and variance are recomputed using mini-batches from the training dataset.
XBAR Simulations and models
PCM models
IBM Zurich team
eleftheriou2019dl_acceleration_on _cm
nandakumar2020mixedprecision_dl_on_cm