CrossSim 3.1

Major Changes in CrossSim 3.1:

• New native Torch and Keras Interfaces for direct conversion of existing networks to use CrossSim for inference and training (Torch only) within existing Torch/Keras flows. See Torch and Keras directories for more details.
• New RRAM Model (RRAM_Wan)
• Assorted bugfixes

Deprecation Notice:

• The previous CrossSim neural network interface (using the DNN class) is deprecated and will be removed in version 3.2. The new Torch and Keras interfaces provide all of the functionality of the previous interface.

CrossSim 3.0.3

• DNN inference interface: added feature to export all array conductances to file for a DNN.

CrossSim 3.0.2

• Fix parasitic handling with different per-row and per-column parasitic resistance (#18).

CrossSim 3.0.1

• Fix AnalogCore parameter initialization issue when using multiple different CrossSimParameters objects
• Fix AnalogCore initialization issue when using cupy

CrossSim 3.0

Major Changes in CrossSim 3.0:

• New AnalogCore interface providing a numpy-like interface for MVM and VMM operations
• Overhauled and reorganized CrossSimParameters to clarify settings and simplify behavior including JSON load/save support
• Neural network training and outer product update support removed If you need those features for your research please continue to use CrossSim2.0
• New ADC, DAC, and device interfaces to simplify adding new models and mixing and matching behaviors
• New circuit-based ADC models
• Optional matrix-multipy fast-path providing up to 40x speedup on certain workloads including CNNs
• Complex number support
• Matlab compatibility
• Assorted modeling improvements

CrossSim 3.0b1

Pre-release of CrossSim 3.0.

Major Changes in CrossSim 3.0:

• New AnalogCore interface providing a numpy-like interface for MVM and VMM operations
• Overhauled and reorganized CrossSimParameters to clarify settings and simplify behavior including JSON load/save support
• Neural network training and outer product update support removed If you need those features for your research please continue to use CrossSim2.0
• New ADC, DAC, and device interfaces to simplify adding new models and mixing and matching behaviors
• New circuit-based ADC models
• Optional matrix-multipy fast-path providing up to 40x speedup on certain workloads including CNNs
• Complex number support
• Matlab compatibility
• Assorted modeling improvements

CrossSim 2.0

CrossSim 2.0 supports neural network inference and training using analog in-memory computing providing data loading and configuration scripts for several common neural networks and datasets which interface with the CrossSim core.

Note: Training support will be removed from subsequent releases until a long-term feature maintainer can be found. If you need training for your research, do not upgrade to later versions.