SimNet Compiled FPGA Models
Fuente:
Zenodo
Salvato in:
| Autore principale: | |
|---|---|
| Natura: | Recurso digital |
| Pubblicazione: |
Zenodo
2026
|
| Soggetti: | |
| Accesso online: | |
| Tags: |
Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
|
| _version_ | 1866901955885924352 |
|---|---|
| author | Marcus Fredriksson |
| author_facet | Marcus Fredriksson |
| contents | <p><strong>Description:</strong> A collection of 25 compiled neural network models (.xmodel format) for the Xilinx ZCU104 FPGA with a B4096 DPU accelerator (DPUCZDX8G ISA 2). The models are members of the SimNet family, which is a set of lightweight 1D convolutional neural networks (1D-CNNs) for real-time anomaly detection on sliding windows of rubber bushing time-series data. Models span five depth tiers (3–7 convolutional stages) and a parameter range from approximately 25 thousand to 510 million parameters.</p> <p><strong>Models:</strong></p> <table> <tbody> <tr> <td>Model</td> <td>Parameters</td> <td>Depth</td> </tr> <tr> <td>model-25k</td> <td>25,385</td> <td>3</td> </tr> <tr> <td>model-39k</td> <td>39,385</td> <td>3</td> </tr> <tr> <td>model-59k</td> <td>58,713 </td> <td>3</td> </tr> <tr> <td>model-85k</td> <td>84,521</td> <td>3</td> </tr> <tr> <td>model-129k</td> <td>128,545</td> <td>3</td> </tr> <tr> <td>model-198k</td> <td>197,705</td> <td>4</td> </tr> <tr> <td>model-282k</td> <td>282,169</td> <td>4</td> </tr> <tr> <td>model-440k</td> <td>439,545</td> <td>4</td> </tr> <tr> <td>model-658k</td> <td>657,921</td> <td>4</td> </tr> <tr> <td>model-973k</td> <td>972,537</td> <td>4</td> </tr> <tr> <td>model-1_5m</td> <td>1,473,537</td> <td>5</td> </tr> <tr> <td>model-2_2m</td> <td>2,229,361</td> <td>5</td> </tr> <tr> <td>model-3_4m</td> <td>3,378,601</td> <td>5</td> </tr> <tr> <td>model-5_1m</td> <td>5,051,921</td> <td>5</td> </tr> <tr> <td>model-7_6m</td> <td>7,583,625</td> <td>5</td> </tr> <tr> <td>model-11m</td> <td>11,398,681</td> <td>6</td> </tr> <tr> <td>model-17m</td> <td>17,283,305</td> <td>6</td> </tr> <tr> <td>model-26m</td> <td>25,840,833</td> <td>6</td> </tr> <tr> <td>model-39m</td> <td>38,802,553</td> <td>6</td> </tr> <tr> <td>model-58m</td> <td>58,406,369</td> <td>6</td> </tr> <tr> <td>model-88m</td> <td>88,211,025</td> <td>7</td> </tr> <tr> <td>model-132m</td> <td>132,307,793</td> <td>7</td> </tr> <tr> <td>model-200m</td> <td>199,565,105</td> <td>7</td> </tr> <tr> <td>model-boundary</td> <td>~300,000,000</td> <td>7</td> </tr> <tr> <td>model-overflow</td> <td>~510,000,000</td> <td> </td> </tr> </tbody> </table> <p>PyTorch training checkpoints were quantised and compiled using pytorch-nndct 3.5.0 (Vitis AI 3.5.0, opset 17, INT8 post-training quantisation) targeting the ZCU104 B4096 DPU fingerprint. The .xmodel files require Vitis AI VART runtime >= 3.0 on aarch64 (<a title="Synthetic Rubber Bushing Time-Series" href="https://doi.org/10.5281/zenodo.19846170" rel="noopener">dataset found here</a>)</p> <p>All models accept a single inference window of shape (1, 512, 3), batch size 1, 512 time steps, 3 channels (x_meas, v_meas, F_meas) scaled by the provided StandardScaler. Output shape is (1,), a single anomaly score (sigmoid logit).</p> |
| format | Recurso digital |
| id | zenodo_https___doi_org_10_5281_zenodo_19847357 |
| institution | Zenodo |
| language | |
| publishDate | 2026 |
| publisher | Zenodo |
| record_format | zenodo |
| spellingShingle | SimNet Compiled FPGA Models Marcus Fredriksson fpga dpu xmodel vitis ai cnn anomaly detection quantization xilinx zcu104 energy-per-inference dpuczdx8g <p><strong>Description:</strong> A collection of 25 compiled neural network models (.xmodel format) for the Xilinx ZCU104 FPGA with a B4096 DPU accelerator (DPUCZDX8G ISA 2). The models are members of the SimNet family, which is a set of lightweight 1D convolutional neural networks (1D-CNNs) for real-time anomaly detection on sliding windows of rubber bushing time-series data. Models span five depth tiers (3–7 convolutional stages) and a parameter range from approximately 25 thousand to 510 million parameters.</p> <p><strong>Models:</strong></p> <table> <tbody> <tr> <td>Model</td> <td>Parameters</td> <td>Depth</td> </tr> <tr> <td>model-25k</td> <td>25,385</td> <td>3</td> </tr> <tr> <td>model-39k</td> <td>39,385</td> <td>3</td> </tr> <tr> <td>model-59k</td> <td>58,713 </td> <td>3</td> </tr> <tr> <td>model-85k</td> <td>84,521</td> <td>3</td> </tr> <tr> <td>model-129k</td> <td>128,545</td> <td>3</td> </tr> <tr> <td>model-198k</td> <td>197,705</td> <td>4</td> </tr> <tr> <td>model-282k</td> <td>282,169</td> <td>4</td> </tr> <tr> <td>model-440k</td> <td>439,545</td> <td>4</td> </tr> <tr> <td>model-658k</td> <td>657,921</td> <td>4</td> </tr> <tr> <td>model-973k</td> <td>972,537</td> <td>4</td> </tr> <tr> <td>model-1_5m</td> <td>1,473,537</td> <td>5</td> </tr> <tr> <td>model-2_2m</td> <td>2,229,361</td> <td>5</td> </tr> <tr> <td>model-3_4m</td> <td>3,378,601</td> <td>5</td> </tr> <tr> <td>model-5_1m</td> <td>5,051,921</td> <td>5</td> </tr> <tr> <td>model-7_6m</td> <td>7,583,625</td> <td>5</td> </tr> <tr> <td>model-11m</td> <td>11,398,681</td> <td>6</td> </tr> <tr> <td>model-17m</td> <td>17,283,305</td> <td>6</td> </tr> <tr> <td>model-26m</td> <td>25,840,833</td> <td>6</td> </tr> <tr> <td>model-39m</td> <td>38,802,553</td> <td>6</td> </tr> <tr> <td>model-58m</td> <td>58,406,369</td> <td>6</td> </tr> <tr> <td>model-88m</td> <td>88,211,025</td> <td>7</td> </tr> <tr> <td>model-132m</td> <td>132,307,793</td> <td>7</td> </tr> <tr> <td>model-200m</td> <td>199,565,105</td> <td>7</td> </tr> <tr> <td>model-boundary</td> <td>~300,000,000</td> <td>7</td> </tr> <tr> <td>model-overflow</td> <td>~510,000,000</td> <td> </td> </tr> </tbody> </table> <p>PyTorch training checkpoints were quantised and compiled using pytorch-nndct 3.5.0 (Vitis AI 3.5.0, opset 17, INT8 post-training quantisation) targeting the ZCU104 B4096 DPU fingerprint. The .xmodel files require Vitis AI VART runtime >= 3.0 on aarch64 (<a title="Synthetic Rubber Bushing Time-Series" href="https://doi.org/10.5281/zenodo.19846170" rel="noopener">dataset found here</a>)</p> <p>All models accept a single inference window of shape (1, 512, 3), batch size 1, 512 time steps, 3 channels (x_meas, v_meas, F_meas) scaled by the provided StandardScaler. Output shape is (1,), a single anomaly score (sigmoid logit).</p> |
| title | SimNet Compiled FPGA Models |
| topic | fpga dpu xmodel vitis ai cnn anomaly detection quantization xilinx zcu104 energy-per-inference dpuczdx8g |
| url | https://doi.org/10.5281/zenodo.19847357 |