KO | EN

GitLite

Minimalist Open Source Repository Search Engine

gitlite — search

#dsh #dsh-plugin #developer-tools #ai-agent #claude-code #codex #typescript #ai-agents #deepseek-harness #mcp #cli #rust

ASiM

★ 30 Open GitHub ↗

[TVLSI 2025] ACiM Inference Simulation Framework in "ASiM: Modeling and Analyzing Inference Accuracy of SRAM-Based Analog CiM Circuits"

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# ASiM

ASiM is an inference simulation framework for SRAM-based ACiM design developed by Keio CSG. We hope it would help space exploration of ACiM circuit design and fast inference evaluation on a prototype chip. For any discussion, please feel free to comment on the "Issues" tab, thanks.

## 1. Overview

### 1.1. Introduction to the Directory

* `./bins`: Some useful scripts to process datasets and model weights.
* `./dataset`: Datasets for CiM benchmarks.
* `./model`: DNN models for simulations.
* `./module`: ASiM modules for Bit-wise simulation.
* `./tools`: Some useful tools.
* `./main`: Main directory for simulations.

### 1.2. Quick Start

All you need to do is to modify the parameter/argument settings `config.py` in the main folder. Then run `src_train.py` for model training and `src_simulation.py` for simulations.

When you run `src_train.py`, the training log and model weight will be generated in `main/run`.

## 2. Environment

We handle this project with the following environment:

```
python==3.11.7
torch==2.3.0
numpy==1.24.3
matplotlib==3.8.0
easydict==1.13
```

## 3. Simulation Time

A personal GPU will be good to handle most conditions. Simulation can be run with CPU for ResNet-18/CIFAR-10 level task. For ViT simulations on ImageNet, we suggest to run on a GPU server.

ASiM can be run with various devices and an estimation of running time (w/ full validation set) is listed as follows:

| Environment           | Batch Size | Dataset  | Model            | Duration (s) |
|-----------------------|------------|----------|------------------|--------------|
| i9-13980HX (64G DDR5) | 128        | CIFAR-10 | ResNet-18 (W4A4) | 6000         |
| RTX4080-Laptop (12G)  | 1024       | CIFAR-10 | ResNet-18 (W4A4) | 230          |
| RTX4080-Laptop (12G)  | 256        | ImageNet | ResNet-18 (W6A6) | 6645         |
| A6000 (48G)           | 512        | ImageNet | ResNet-18 (W6A6) | 4049         |
| A6000 (48G)           | 1024       | ImageNet | ViT-B-32 (W8A8)  | 19841        |

## 4. ASiM

### 4.1. Abstract

<a href="https://arxiv.org/abs/2411.11022" target="_blank">Paper Link</a>

SRAM-based Analog Compute-in-Memory (ACiM) demonstrates promising energy efficiency for deep neural network (DNN) processing. Nevertheless, efforts to optimize efficiency frequently compromise accuracy, and this trade-off remains insufficiently studied due to the difficulty of performing full-system validation. Specifically, existing simulation tools rarely target SRAM-based ACiM and exhibit inconsistent accuracy predictions, highlighting the need for a standardized, SRAM CiM circuit-aware evaluation methodology. This paper presents ASiM, a simulation framework for evaluating inference accuracy in SRAM-based ACiM systems. ASiM captures critical effects in SRAM based analog compute in memory systems, such as ADC quantization, bit parallel encoding, and analog noise, which must be modeled with high fidelity due to their distinct behavior in charge domain architectures compared to other memory technologies. ASiM supports a wide range of modern DNN workloads, including CNN and Transformer-based models such as ViT, and scales to large-scale tasks like ImageNet classification. Our results indicate that bit-parallel encoding can improve energy efficiency with only modest accuracy degradation; however, even 1 LSB of analog noise can significantly impair inference performance, particularly in complex tasks such as ImageNet. To address this, we explore hybrid analog–digital execution and majority voting schemes, both of which enhance robustness without negating energy savings. ASiM bridges the gap between hardware design and inference performance, offering actionable insights for energy-efficient, high-accuracy ACiM deployment.

### 4.2. Mode Setting for ASiM Modules
Users can import and replace ASiM modules with default nn.Module into their own DNN models. ASiM modules have additional arguments in module initialization, and all configurable arguments are extracted into `config.py` within this framework. The meaning of each argument can be found in this configuration file. ASiM have two main modes in general use: **Train** and **Simulation**, for model training/fine-tuning and running bit-wise simulation.

#### 4.2.1. Running Simulations by ASiM

For running simulation, a pre-trained model weight has to be specified in `config.py`: `cfg.pretrain_model_path = 'model_path.pkl'`.

The following example shows an ASiM module configuration for simulation:

```
ASiMConv2d(in_planes=64,
           out_planes=128,
           kernel_size=3,
           stride=1,
           padding=1,
           bias=False,
           wbit=8,
           xbit=8,
           adc_prec=8,
           nrow=256,
           rand_noise_sigma=0.05,
           non_linear_sigma=0.05,
           act_enc=None,
           signed_act=False,
           hybrid_levels=None,
           mode='Simulation',
           trim_noise=0.0,
           device='cuda')
```

Please note that the mode must be set as `mode='Simulation'` to switch ASiM modules into simulation mode.

#### 4.2.2. Training DNN Models by ASiM

The following example shows an ASiM module configuration for model training, with mode set as `mode='Train'`:

```
ASiMConv2d(in_planes=64,
           out_planes=128,
           kernel_size=3,
           stride=1,
           padding=1,
           bias=False,
           wbit=8,
           xbit=8,
           adc_prec=8,
           nrow=256,
           rand_noise_sigma=0.0,
           non_linear_sigma=0.0,
           act_enc=None,
           signed_act=False,
           hybrid_levels=None,
           mode='Train',
           trim_noise=0.0,
           device='cuda')
```

The `wbit` and `xbit` have to be specified for quantization-aware training. For training tiny scale DNN models (ResNet-18) with simple tasks (CIFAR-10/CIFAR-100), users can directly train the model quickly. For relately larger models (ViT) and ImageNet task, we recommend to download and process PyTorch official model weight for fine-tuning, with pretrained model weight directory specified as `cfg.pretrain_model_path = 'model_path.pkl'`.

#### 4.2.3. Fine-tuning a Robust DNN Model for ACiM Applications

To perform noise-aware training for obtaining a more robust DNN model, users can configure ASiM modules like:

Here, a pretrained model weight must be given as `cfg.pretrain_model_path = 'model_path.pkl'`, and training noise intensity must be given like `trim_noise=50.0`. After the noise-aware training, the DNN model can tolerate a certain level of analog noise.

### 4.3. Download Our Post-NAT-Models

| File Name                                | Dataset  | Model            | Link                                                                                                                                     |
|------------------------------------------|----------|------------------|------------------------------------------------------------------------------------------------------------------------------------------|
| resnet18_cifar10_w8a8_pact_trim_100.pkl  | CIFAR-10 | ResNet-18 (W8A8) | <a href="https://drive.google.com/file/d/1FoVjgLTTTZhFwD7plXZv1eWMHZzpEa-p/view?usp=drive_link" target="_blank">Google Drive</a>         |
| resnet18_imagenet_w8a8_pact_trim_50.pkl  | ImageNet | ResNet-18 (W8A8) | <a href="https://drive.google.com/file/d/1G42fgOUXJejV0f6rDMj2k3gDq1r7b3Z0/view?usp=drive_link" target="_blank">Google Drive</a>         |
| vitb32_cifar10_w8a8_trim_60.pkl          | CIFAR-10 | ViT-B-32 (W8A8)  | <a href="https://drive.google.com/file/d/1uypqLxoZA0hBX-UUOpk1-HsPAb3GDUqJ/view?usp=drive_link" target="_blank">Google Drive</a>         |
| vitb32_imagenet_w8a8_trim_40.pkl         | ImageNet | ViT-B-32 (W8A8)  | <a href="https://drive.google.com/file/d/1-3togKT0lIMjk4Bhsb5jO_DoqaNbUhGI/view?usp=drive_link" target="_blank">Google Drive</a>         |

## 5. Guiding Setup by Reproducing Results in the Paper

In this section, we will give the setup guide by reproducing some results reported in the paper. Please follow the steps and familiarize how ASiM works.

### 5.1. Prepare Environment and Datasets

Please refer our software versions provided in Sec. 2 for environment settings. Install easydict if you do not have it, and download our model weights provided in Sec. 4.3.

Download the following datasets: <a href="https://drive.google.com/file/d/1b7auKQzYE0VTe_0JCGCo8S6SzEl-WvuE/view?usp=drive_link" target="_blank">CIFAR-10</a>, <a href="https://drive.google.com/file/d/143wrVmpfqlH1Y0URvFVTFgyVhicO3ZA5/view?usp=drive_link" target="_blank">CIFAR-100</a>, and <a href="https://www.kaggle.com/c/imagenet-object-localization-challenge/overview/description" target="_blank">ImageNet</a>, respectively. Please use `./bins/valid_setup.py` to reformat the validation folder to match our dataset file of `./dataset/imagenet.py`.

The recollected dataset path should align with the following form:

CIFAR-10:

CIFAR-100:

ImageNet:

```
|-- ImageNet
|   |-- train
|   |   |-- n01440764
|   |   |-- ...
|   |-- valid
|   |   |-- n01440764
|   |   |-- ...
```

### 5.2. Simulation Examples

#### 5.2.1. ResNet-18 on CIFAR-10

For simulations of ResNet-18 on CIFAR-10, please refer configuration at `./examples/resnet18_cifar10/config.py`. Please give the dataset path of validation set to `cfg.test_dir`, and pre-trained model weight to `cfg.pretrain_model_path`. Based on your devices, please set the appropriate batch size and work loader to `cfg.test_bs` and `cfg.test_workers`.

The default configuration is based on an 8b/8b model simulation on a bit-serial ACiM macro with row-parallelsim of 255. The CONV and Linear layer are configured separately.

The model quantization bit can be modified by `wbit` and `xbit` parameters. The row-parallelsim and ADC precision can be modified by `adc_prec` and `nrow` parameters. For analog noise injection, please give `rand_noise_sigma` and `non_linear_sigma` based on your Silicon measurement. For bit-parallel scheme simulation, configure `act_enc` parameter with the encoding bit width (2, 4, etc.). To verify the efficacy of HCiM, configure the `cfg.asim_cnn_hybrid_levels` parameter with hybrid boundary level (1, 2, 3, etc.) to prevent MSB cycles from noise impact.

#### 5.2.2. ResNet-18 on ImageNet

Similarly, please set dataset/model weight path and refer `./examples/resnet18_imagenet/config.py` for configurations on ImageNet.

#### 5.2.3. ViT-B-32 on CIFAR-10

In Transformer models, the configuration of MultiheadAttention, MLP, and FC layers are given separately. Notably, in MultiheadAttention layer, we further give individual settings to QK, AV, and WQKV/WO projection, respectively. Please note the Keys (K) and Activations in projection layers are have a sign bit that have to be processed individually in bit-serial manner, we set `attn_qk_kbit` and `attn_proj_xbit` to 9 so that the LSBs can be evenly divided by 2 and 4, respectively. The HCiM boundary level in ViT is in `cfg.asim_vit_hybrid_levels`. Please refer `./examples/vitb32_cifar10/config.py` for more details.

#### 5.2.4. ViT-B-32 on ImageNet

Similarly, please set dataset/model weight path and refer `./examples/vitb32_imagenet/config.py` for configurations on ImageNet.

#### 5.2.5. Run Simulations

After the above configuration settings, just run `./main/src_simulation.py` for ASiM simulations.

### 5.3. Training Examples

We give an example of training ResNet-18 model on CIFAR-10 dataset for ACiM applications.

#### 5.3.1. Quantization-Aware Training (QAT)

For model training, the `cfg.train_dir` and `cfg.valid_dir` have to be set as your dataset root. To perform training without an pre-trained model, just give `cfg.pretrain_model_path` a **None**. If you wish to start by a pre-trained state, please give the directory of pre-trained model weight. Please note that you may have to rearrange the parameters from original PyTorch provided models, in these cases, the state_dict have to be reformed (refer scripts in `./bins`).

Please set batch size and load workers by `cfg.train_bs`, `cfg.valid_bs`,  and `cfg.train_workers` respectively. Also, the `wbit` and `xbit` need to be configured to quantization weight and activation bit width, respectively.

#### 5.3.2. Noise-Aware Training (NAT)

For NAT, please give a pre-trained model to `cfg.pretrain_model_path`, and configure the noise intensity to `trim_noise` parameter.

#### 5.3.3. Train Models

After the above configuration settings, just run `./main/src_train.py` for model training.

Kindly remind that training conditions such as epoch/learning rate may need to be configured case by case (`cfg.max_epoch` and `cfg.lr_init`). A `run` directory will be generated when the training is finished that containing all training log and model weights.

## 6. Model Extension

As ACiM technology rapidly evolves, novel designs continue to enhance area efficiency, energy efficiency, and throughput. Although ASiM is primarily developed for charge domain SRAM ACiM, we encourage users to extend the model by tailoring the code in ASiM modules to reflect their novel circuit designs, ensuring alignment with their specific hardware characteristics. Here, we give two additional examples using ASiM as baseline to show the model's flexibility.

### 6.1. RRAM Crossbar

In the case of RRAM, where the crossbar cells encode multi-bit weights directly, the loop over binary weight bits in the original simulation algorithm is unnecessary and can be omitted. Additionally, RRAM circuits are primarily affected by conductance variation, making the original capacitor nonlinearity model for SRAM-based ACiM unsuitable. Instead, we model conductance drift as a random perturbation characterized by a percentage-based standard deviation. With these minor adjustments, the ASiM framework can effectively simulate RRAM-based architectures with satisfactory accuracy. The implementation example of RRAM crossbar ACIM can be found in `./extensions/rram`.

### 6.2. Full Weight Bit-Parallel

By adjusting the order of ideal MAC computing and noise injection & ADC quantization, the ASiM can be extended to further support weight bit-parallel scheme, which is a more aggressive strategy to boost energy efficiency. The implementation example of an full bit-parallel 4b/4b CNN can be found in `./extensions/w4a4_weight_full_bit_parallel`.

### 6.3. Current Domain SRAM ACiM

For current domain SRAM ACiM, replacing capacitance mismatch based non-linearity model with gain compression can effectively capture the non-linearity introduced by current roll-off when transistors transition from saturation to the triode region.

MAC_REAL = MAC_IDEAL / (1 + Alpha * MAC_IDEAL)

Here, MAC_IDEAL represents the ideal MAC value, while MAC_REAL denotes the actual computed value. In current domain SRAM ACiM, the discharge current of the bit-line or sampling capacitor depends on the number of active transistors where the local dot product is ‘1’. When a large number of cell transistors simultaneously discharge the bit-line or capacitor, the rapid drop in Vds shifts the operation into the triode region, leading to current roll-off and introducing non-linearity into MAC computation. To replicate this behavior in simulations, we introduce an empirical parameter Alpha to align the non-linearity. By calibrating Alpha, the real MAC value remains nearly unchanged for small MAC values, while for larger MAC values, the real MAC output decreases due to gain compression effects. Additionally, incorporating random variations into binary weight tensors allows ASiM to accurately model transistor variation in current domain computations. The implementation example of current domain SRAM ACIM in CNN can be found in `./extensions/current_domain`.

## Citation

If you find this repo is useful, please cite our paper. Thanks.

```bibtex
@article{zhang2025asim,
  title={ASiM: Modeling and Analyzing Inference Accuracy of SRAM-Based Analog CiM Circuits},
  author={Zhang, Wenlun and Ando, Shimpei and Chen, Yung-Chin and Yoshioka, Kentaro},
  journal={IEEE Transactions on Very Large Scale Integration (VLSI) Systems},
  number={01},
  pages={1--14},
  year={2025},
  publisher={IEEE Computer Society}
}
```

ASiM

Explore Similar Repositories

Was this content helpful?

Recent Feedback

Download README