IBM Analog Hardware Accelerator Kit 1.1.0

[1.1.0] - 2026/02/03

Added

• Add newly uploaded resources for CPU-only wheels (#739)
• Add a new drift compensation mechanism which uses an ideal reference readout. In the default global drift compensation mechanism, all non-idealities (as set by the corresponding rpu_config) are modeled, potentially resulting in sub-optimal drift compensation scales being computed in some scenarios, e.g., where the output noise is sufficiently large.
• New example (examples/36_gpt2_on_wikitext.py) to run Distil GPT2 on aihwkit with text prediction. (#754)

Removed/Deprecated

• Function convert_to_analog_mapped in src/aihwkit/nn/conversion.py deprecated and removed.

Changed

• Replace legacy release-build workflow with the updated build process (#744)
• Point the online demo link to the correct destination (#743)
• Update bundled notebook wheel to the GPU-enabled 1.0.0 release (#741)
• Modifications/Cleanup/Improvements to Time-Stepped Circuit IRDrop example and model implementation (#753)

Fixed

• Fix memory issues and bugs in analog training for CUDA (#732)
• Fix convert_to_analog for conv layers (#627)

IBM Analog Hardware Accelerator Kit 1.0.0

[1.0.0] - 2025/05/19

Added

• Add new Weight Programming Optimization Feature (#703)
• Add Floating Point Preset for inference (#705)
• Add new notebook on a tutorial for Analog Device Non Idealities setup (#682)
• Add new simulation for Analog Filamentary Conductive-Metal-Oxide (CMO)/HfOx ReRAM device noise models for inference (#702)
• Add new Quantization library (#719)
• Add new Half Precision training section on Using the Simulator documentation and related example (#678)

Fixed

• Fix Hardware-Aware training tutorial notebooks (#700)
• Fix Post-Training Input Range Calibration notebook (#716)

Changed

• Change License from Apache 2.0 to MIT (#696)
• Update examples index for better navigation (#704)
• Update tutorial notebooks to run with latest builds of aihwkit (#713)
• Update and enhance API documentation on IOParameters for is_perfect (#718)
• Migrate from Travis CI/CD integration to Github Actions (#720)

IBM Analog Hardware Accelerator Kit 0.9.2

[0.9.2] - 2024/09/18

Added

• Added new Hermes noise model and related notebooks (#685)
• Added new conductance converters (#685)
• Make Conv layers also compatible with non-batched inputs (#685)
• Added per column drift compensation (#685)
• Added custom drifts (#685)

Changed

• Update requirements-examples.txt (#685)

IBM Analog Hardware Accelerator Kit 0.9.1

[0.9.1] - 2024/05/16

Added

• Added column wise scaling logic to fusion import/export to improve accuracy (#652)
• Added a new example that demonstrates how to import and perform inference using a model which has been trained in a hardware-aware fashion using an external library (#648)
• Added a wew WeightModifierType.DISCRETIZE_PER_CHANNEL type and a test case to validate the correctness against manual quantization in PyTorch (#618)

Fixed

• Fix sub-optimal mapping of conductances to weights for fusion by regressing weights per column (#653)
• Documentation correction: Use ADDITIVE_CONSTANT instead of ADD_NORMAL` in WeightNoiseType (#630)
• Fix continuing training based on checkpoint using torch tile (#626)
• Fix the support of different dtypes for the torch model (#625)
• Fixes the fall-through to the default error message when using drop connect (#624)
• Update analog_fusion notebook (#611)

IBM Analog Hardware Accelerator Kit 0.9.0

[0.9.0] - 2024/01/26

Added

• On-the-fly change of some RPUConfig fields (#539)
• Fusion chip CSV file model weights exporter functionality (#538)
• Experimental support for RPU data types (#563)
• Optional AIHWKIT C++ extension module (#563)
• Variable mantissa / exponent tensor conversion operator (#563)
• To digital feature for analog layers (#563)
• New PCM_NOISE type for hardware-aware training for inference (#563)
• Transfer compounds using torch implementation (TorchTransferTile) (#567)
• Weight programming error plotting utility (#572)
• Add optimizer checkpoint in example 20 (#573)
• Inference tile with time-dependent IR-drop (#587)
• Linear algebra module (#588)
• New Jupyter notebook for Fusion chip access (#601)

Fixed

• Repeated call of cuda() reset the weights for InferenceTile (#540)
• Custom tile bugfixes (#563)
• Bug-fixes for specialized learning algorithms (#563)
• Bug-fix for data-parallel hardware-aware training for inference (#569)
• Fix docker build stubgen (#581)
• Fix readthedoc builds (#586)
• Fix the backward of the input ranges in the torch tile (#606)

Changed

• Parameter structure changed into separate files to reduce file sizes (#563)
• RPUConfig has a new runtime field and inherits from additional base classes (#563)
• AnalogWrapper now directly adds module classes to subclasses (#563)
• RNN linear layers more custonable (#563)
• Parameters for specialized learning algorithms changed somwhat (#563)
• RNN modules inherit from Module or AnalogContainerBase instead of AnalogSequential (#563)
• Adjustment of parameter to bindings for various number formats (#563)
• Documentation updates and fixes (#562, #564, #570, #575, #576, #580, #585, #586)
• Updated installation instructions in Readthedoc (#594)

IBM Analog Hardware Acceleration Kit 0.8.0

[0.8.0] - 2023/07/14

Added

• Added new tutorial notebooks to cover the concepts of training,
  hardware-aware training, post-training calibration, and extending aihwkit functionality (#518, #523, #526)
• Calibration of input ranges for inference (#512)
• New analog in-memory training algorithms: Chopped Tiki-taka II (#512)
• New analog in-memory training algorithms: AGAD (#512)
• New training presets: ReRamArrayOMPresetDevice,
  ReRamArrayHfO2PresetDevice, ChoppedTTv2*, AGAD* (#512)
• New correlation detection example for comparing specialized analog SGD
  algorithms (#512)
• Simplified build_rpu_config script for generating RPUConfigs for
  analog in-memory SGD (#512)
• CustomTile for customization of in-memory training algorithms (#512)
• Pulse counters for pulsed analog training (#512)
• TorchInferenceTile for a fully torch-based analog tile for
  inference (not using the C++ RPUCuda engine), supporting a subset of MVM nonidealities (#512)
• New inference preset StandardHWATrainingPreset (#512)
• New inference noise model ReRamWan2022NoiseModel (#512)
• Improved HWA-training for inference featuring input and output range
  learning and more (#512)
• Improved CUDA memory management (using torch cached GPU memory for
  internal RPUCuda buffer) (#512)
• New layer generator: analog_layers() loops over layer modules (except
  container) (#512)
• AnalogWrapper for wrapping a full torch module (Without using
  AnalogSequential) (#512)
• convert_to_digital utility (# 512)
• TileModuleArray for logical weight matrices larger than a single tile. (#512)
• Dumping of all C++ fields for accurate analog training saving and
  training continuation after checkpoint load. (#512)
• apply_write_noise_on_set for pulsed devices. (#512)
• Reset device now also for simple devices. (#512)
• SoftBoundsReference, PowStepReference for explicit reference
  subtraction of symmetry point in Tiki-taka (#512)
• Analog MVM with output-to-output std-deviation variability
  (output_noise_std) (#512)
• Plotting utility for weight errors (#512)
• per_batch_sample weight noise injections for TorchInferenceRPUConfig (#512)

Fixed

• BERT example 24 using AnalogWrapper (#514)
• Cuda supported testing in examples based on AIHWKIT compilation (#513)
• Fixed compilation error for CUDA 12.1. (#500)
• Realistic read weights could have applied the scales wrongly (#512)

Changed

• Major re-organization of AnalogTiles for increased modularity
  (TileWithPeriphery, SimulatorTile, SimulatorTileWrapper). Analog
  tile modules (possibly array of analog tiles) are now also torch Module. (#512)
• Change in tile generators: analog_model.analog_tiles() now loops over
  all available tiles (in all modules) (#512)
• Import and file position changes. However, user can still import RPUConfig
  related modules from aihwkit.simulator.config (#512)
• convert_to_analog now also considered mapping. Set
  mapping.max_out_size = 0 and mapping.max_out_size = 0 to avoid this. (#512)
• Mapped layers now use TileModuleArray array by default. (#512)
• Checkpoint structure is different than previous
  versions. utils.legacy_load provides a way to load old checkpoints. (#512)

Removed

• realistic_read_write is removed from some high-level function. Use
  program_weights (after setting the weights) or read_weights
  for realistic reading (using weight estimation technique). (#512)

IBM Analog Hardware Acceleration Kit 0.7.1

Added

• Updated the CLI Cloud runner code to support inference experiment result. (#491)
• Read weights is done with least-square estimation method. (#489)

Fixed

• Realistic read / write behavior was broken for some tiles. (#489)

Changed

• Torch minimal version has changed to version 1.9. (#489)
• Realistic read / write is now achieved by read_weights and program_weights. (#489)

Removed

• The tile methods get/set_weights_realistic are removed. (#489)

IBM Analog Hardware Acceleration Kit 0.7.0

[0.7.0] - 2023/01/30

Added

• Reset tiles method (#456)
• Added many new analog MAC non-linearties (forward / backward pass) (#456)
• Polynomial weight noise for hardware-aware training (#456)
• Remap functionality for hardware-aware training (#456)
• Input range estimation for InferenceRPUConfig (#456)
• CUDA always syncs and added non-blocking option if not wished (#456)
• Fitting utility for fitting any device model to conductance measurements (#456)
• Added PowStepReferenceDevice for easy subtraction of symmetry
  point (#456)
• Added SoftBoundsReferenceDevice for easy subtraction of symmetry
  point (#456)
• Added stand-alone functions for applying inference drift to any model (#419)
• Added Example 24: analog inference and hardware-aware training on BERT with the SQUAD task (#440)
• Added Example 23: how to use AnalogTile directly to implement an
  analog matrix-vector product without using pytorch modules. (#393)
• Added Example 22: 2 layer LSTM network trained on War and Peace dataset. (#391)
• Added a new notebook for exploring analog sensitivities. (#380)
• Remapping functionality for InferenceRPUConfig. (#388)
• Inference cloud experiment and runners. (#410)
• Added analog_modules generator in AnalogSequential. (#410)
• Added SKIP_CUDA_TESTS to manually switch off the CUDA tests.
• Enabling comparisons of RPUConfig instances. (#410)
• Specific user-defined function for layer-wise setting for RPUConfigs
  in conversions. (#412)
• Added stochastic rounding options for MixedPrecisionCompound. (#418)
• New remap parameter field and functionality in
  InferenceRPUConfig (#423).
• Tile-level weight getter and setter have apply_weight_scaling
  argument. (#423)
• Pre and post-update / backward / forward methods in BaseTile for
  easier user-defined modification of pre and/or post-processings of a tile. (#423)
• Type-checking for RPUConfig fields. (#424)

Fixed

• Decay fix for compound devices (#463)
• RPUCuda backend update with many fixes (#456)
• Missing zero-grad call in example 02 (#446)
• Indexing error in OneSidedDevice for CPU (#447)
• Analog summary error when model is on cuda device. (#392)
• Index error when loading the state dict with a model use previously. (#387)
• Weights that were not contiguous could have been set wrongly. (#388)
• Programming noise would not be applied if drift compensation was not
  used. (#389)
• Loading a new model state dict for inference does not overwrite the noise
  model setting. (#410)
• Avoid AnalogContext copying of self pointers. (#410)
• Fix issue that drift compensation is not applied to conv-layers. (#412)
• Fix issue that noise modifiers are not applied to conv-layers. (#412)
• The CPU AnalogConv2d layer now uses unfolded convolutions instead of
  indexed covolutions (that are efficient only for GPUs). (#415)
• Fix issue that write noise hidden weights are not transferred to
  pytorch when using get_hidden_parameters in case of CUDA. (#417)
• Learning rate scaling due to output scales. (#423)
• WeightModifiers of the InferenceRPUConfig are no longer called
  in the forward pass, but instead in the post_update_step
  method to avoid issues with repeated forward calls. (#423)
• Fix training learn_out_scales issue after checkpoint load. (#434)

Changed

• Pylint / mypy / pycodestyle / protobuf version bump (#456)
• All configs related classes can now be imported from
  aihwkit.simulator.config (#456)
• Weight noise visualization now shows the programming noise and drift
  noise differences. (#389)
• Concatenate the gradients before applying to the tile update
  function (some speedup for CUDA expected). (#390)
• Drift compensation uses eye instead of ones for readout. (#412)
• weight_scaling_omega_columnwise parameter in MappingParameter is now called
  weight_scaling_columnwise. (#423)
• Tile-level weight getter and setter now use Tensors instead of numpy
  arrays. (#423)
• Output scaling and mapping scales are now distiniguished, only the
  former is learnable. (#423)
• Renamed learn_out_scaling_alpha parameter in MappingParameter to
  learn_out_scaling and columnwise learning has a separate switch
  out_scaling_columnwise. (#423)

Deprecated

• Input weight_scaling_omega argument in analog layers is deprecated. (#423)

Removed

• The _scaled versions of the weight getter and setter methods are
  removed. (#423)

IBM Analog Hardware Acceleration Kit 0.6.0

[0.6.0] - 2022/05/16

Added

• Set weights can be used to re-apply the weight scaling omega. (#360)
• Out scaling factors can be learnt even if weight scaling omega was set to 0. (#360)
• Reverse up / down option for LinearStepDevice. (#361)
• Generic Analog RNN classes (LSTM, RNN, GRU) uni or bidirectional. (#358)
• Added new PiecewiseStepDevice where the update-step response function can be arbitrarily defined by the user in a piece-wise linear manner. It can be conveniently used to fit any experimental device data. (#356)
• Several enhancements to the public documentations: added a new section for hw-aware training, refreshed the reference API doc, and added the newly supported LSTM layers and the mapped conv layers. (#374)

Fixed

• Legacy checkpoint load with alpha scaling. (#360)
• Re-application of weight scaling omega when loading checkpoints. (#360)
• Write noise was not correctly applied for CUDA if dw_min_std=0. (#356)

Changed

• The set_alpha_scale and get_alpha_scale methods of the C++ tiles are removed. (#360)
• The lowest supported Python version is now 3.7, as 3.6 has reached end-of-life. Additionally, the library now officially supports Python 3.10. (#368)

IBM Analog Hardware Acceleration Kit 0.5.1

Added

• Load model state dict into a new model with modified RPUConfig. (#276)
• Visualization for noise models for analog inference hardware simulation. (#278)
• State independent inference noise model. (# 284)
• Transfer LR parameter for MixedPrecisionCompound. (#283)
• The bias term can now be handled either by the analog or digital domain by controlling the digital_bias layer parameter. (#307)
• PCM short-term weight noise. (#312)
• IR-drop simulation across columns during analog mat-vec. (#312)
• Transposed-read for TransferCompound. (#312)
• BufferedTranferCompound and TTv2 presets. (#318)
• Stochastic rounding for MixedPrecisionCompound. (#318)
• Decay with arbitrary decay point (to reset bias). (#319)
• Linear layer AnalogLinearMapped which maps a large weight matrix onto multiple analog tiles. (#302)
• Convolution layers AnalogConvNdMapped which maps large weight matrix onto multiple tiles if necessary. (#331)
• In the new mapping field of RPUConfig the max tile input and output sizes can be configured for the *Mapped layers. (#331)
• Notebooks directory with several notebook examples (#333, #334)
• Analog information summary function. (#302)
• The alpha weight scaling factor can now be defined as learnable parameter by switching learn_out_scaling_alpha in the rpu_config.mapping parameters. (#353)

Fixed

• Removed GPU warning during destruction when using multiple GPUs. (#277)
• Fixed issue in transfer counter for mixed precision in case of GPU. (#283)
• Map location keyword for load / save observed. (#293)
• Fixed issue with CUDA buffer allocation when batch size changed. (#294)
• Fixed missing load statedict for AnalogSequential. (#295)
• Fixed issue with hierarchical hidden parameter settings. (#313)
• Fixed serious issue that loaded model would not update analog gradients. (#302)
• Fixed cuda import in examples. (#320)

Changed

• The inference noise models are now located in aihwkit.inference. (#281)
• Analog state dict structure `has changed (shared weight are not saved). (#293)
• Some of the parameter names of the TransferCompound have changed. (#312)
• New fast learning rate parameter for TransferCompound, SGD learning rate then is applied on the slow matrix (#312).
• The fixed_value of WeightClipParameter is now applied for all clipping types if set larger than zero. (#318)
• The use of generators for analog tiles of an AnalogModuleBase. (#302)
• Digital bias is now accessible through MappingParameter. (#331)
• The aihwkit documentation. New content around analog AI concepts, training presets, analog AI optimizers, new references, and examples. (#348)
• The weight_scaling_omega can now be defined in the rpu_config.mapping. (#353)

Deprecated

• The module aihwkit.simulator.noise_models has been depreciated in favor of aihwkit.inference. (#281)