The AMD Deep Learning Processor Unit (DPU) is a programmable engine primarily designed for accelerating convolutional neural networks (CNNs).
Autoencoders and recurrent neural networks (RNNs) have different levels of support compared to standard CNN models.
Please refer to the following resources to understand DPU support for RNNs:
https://static.eetrend.com/files/2021-12/wen_zhang_/100556407-231958-ug1563-vitis-ai-rnn.pdf
https://github.com/Xilinx/Vitis-AI/tree/2.5/examples/DPU-for-RNN
https://docs.amd.com/r/en-US/ug1414-vitis-ai/Supported-Operators-and-DPU-Limitations
https://docs.amd.com/r/1.4.1-English/ug1414-vitis-ai/Compiling-for-DPU
https://medium.com/data-science/boost-any-machine-learning-model-with-onnx-conversion-de34e1a38266
It is possible to develop a custom kernel that leverages the Versal AI Edge SOM’s AI Engines to accelerate Autoencoder- and RNN-based models, as discussed in the references above. For further details, please check with AMD.
Processing time-series data using Autoencoders and RNNs (including LSTM and GRU) is feasible on AMD Versal AI Edge hardware. This can be achieved by utilizing its heterogeneous architecture, which combines AI Engines-ML for inference acceleration, Programmable Logic (PL) for custom data preprocessing, and Scalar Engines (Arm cores) for system control.
Please note that AMD has discontinued DPU support for Versal AI, and this configuration has not been validated by iWave.
For further product inquiries, please get in touch with our sales team at mktg@iwave-global.com
Please find the detailed product information here:
Versal™ AI Edge System on Module