KO | EN

GitLite

Minimalist Open Source Repository Search Engine

gitlite — search

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

Prosperity

★ 42 Open GitHub ↗

The official implementation of HPCA 2025 paper, Prosperity: Accelerating Spiking Neural Networks via Product Sparsity

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# Prosperity: Accelerating Spiking Neural Networks via Product Sparsity

This repository is the official implementation of HPCA 2025 paper, Prosperity: Accelerating Spiking Neural Networks via Product Sparsity. It contains the cycle-accurate simulator for Prosperity architecture and implementation of several baseline accelerators.

## TL;DR
We propose a Spiking Neural Networks (SNNs) accelerator, **Prosperity**, that leverage a novel **product sparsity** paradigm to improve various SNNs efficiency.

## Usage

### Recommend Environment

- Python == 3.10
- GCC/G++ >= 11.3.0
- CUDA >= 12.1

We recommend using conda to manage the environment.

### Setup

Clone the repository and install the requirements, cloning may take a while due to the large size of the data in the repository.

```bash
conda create -n prosperity python=3.10 -y
conda activate prosperity
pip install -r requirements.txt
git submodule update --init
```

Install Prosperity CUDA kernel
```bash
cd kernels
python setup.py install
cd ..
```

Install CACTI
```bash
cd cacti
make
cd ..
```

### Repository Structure

- `simulator`: Cycle-accurate simulator for Prosperity and baselines
- `kernels`: CUDA kernel for fast simulation of Prosperity
- `data`: The activation matrix data for simulation
- `cacti`: CACTI for buffer area evaluation
- `reference`: Reference results for Prosperity and baselines

### Run Prosperity and baselines simulation (Figure 8) (15 minutes)

```bash
cd simulator
./run_simulator_all.sh
```

This script will run the simulation of Prosperity and baselines on all of the models and datasets shown in Figure 8 in the paper. 
It is expected to take 15 minutes to finish with CUDA.
It will output the end to end runtime and energy consumption of each accelerator and on each dataset.
The results will be stored in `output` folder, named as `time.csv` and `energy.csv`.

We provide several reference files in `reference` folder. [`time_reference.xlsx`](reference/time_reference.xlsx) and [`energy_reference.xlsx`](reference/energy_reference.xlsx) contain reference results and figure 8 in the paper.

The energy evaluation depend on the file `mem_reference.csv`. This stats is derived from the Prosperity simulation output in file `Prosperity_ST_SCNN_256_16_cuda.txt` in `output` folder.

### Design space exploration (Figure 7, without power and area stats) (1 hour 30 minutes)

For two key design choices in Prosperity, tile size on M dimension and K dimension, we provide a script that test different tile_size_M and tile_size_K configuration in Prosperity and organize this result into file `M_dse.csv` for M dimension exploration and file `K_dse.csv` for K dimension exploration.

File [`M_dse_reference.xlsx`](reference/M_dse_reference.xlsx) and [`K_dse_reference.xlsx`](reference/K_dse_reference.xlsx) contain reference DSE results and figure 7 in the paper.

Notice that this script does not include power and area stats, since it is evaluated by synopsys design compiler, we are not able to release the script for this part under the license.

```bash
./run_DSE.sh
```

### Run Sparsity Analysis (Figure 11) (7 minutes)

```bash
python simulator.py --type Prosperity --sparse_analysis_mode --use_cuda
```

The script record the bit density and product density of Prosperity on all the models and datasets in Figure 11 in the paper.
The results will be stored in file `density_analysis.csv`.

The reference file is [`density_analysis_reference.xlsx`](reference/density_analysis_reference.xlsx).
Notice that the output file does not contain the density stats for FS neuron since this result is directly derived from the baseline paper.

### Run Buffer Area Evaluation (Figure 10 buffer area stats) (1 second)

```bash
python buffer_cacti.py
```

This program call the CACTI to evaluate the buffer power, area of Prosperity. It also evaluate the power of DRAM.
The output result should be identical to the buffer area value in left pie chart in Figure 10 in the paper.

### Run Prosperity Ablation Study (Figure 9)

- Black pillar: Eyeriss with 128 PE, scale the time result of following ablation study by 128/168 to get the result used in Figure 9.
```bash
python simulator.py --type Eyeriss
```

- For the orange, green, purple, and blue pillars

```bash
python simulator.py --type PTB
python simulator.py --type Prosperity --bit_sparsity
python simulator.py --type Prosperity --issue_type 1 --use_cuda
python simulator.py --type Prosperity --use_cuda
```

### Applying ProSparsity to LoAS

Our proposed ProSparsity can be applied to other SNN accelerators. We provide an example of applying ProSparsity to LoAS.

```bash
python simulator.py --type LoAS
```

## Citation
If you find Prosperity helpful in your project or research, please consider citing our [paper](https://arxiv.org/abs/2503.03379):
```
@inproceedings{wei2025prosperity,
  title={Prosperity: Accelerating Spiking Neural Networks via Product Sparsity},
  author={Wei, Chiyue and Guo, Cong and Cheng, Feng and Li, Shiyu and Yang, Hao Frank and Li, Hai Helen and Chen, Yiran},
  booktitle={2025 IEEE International Symposium on High Performance Computer Architecture (HPCA)},
  pages={806--820},
  year={2025},
  organization={IEEE}
}
```

## Contributing

Welcome to use the code or contribute to the project!

Prosperity

Explore Similar Repositories

Was this content helpful?

Recent Feedback

Download README