Our Mission for embedded AI
In our mission to unleash AI on the edge, Datakalab has applied its proprietary quantization method on classical model architectures and converted everything to run in .rtm format for seamless compatibility with NXP® hardware, such as I.MX 8M Plus® and others.
Download any of the quantised models below to put your deep learning models on NXP® hardware. Enjoy and let us know what you think!
If you are looking for further compression methods (continual learning, pruning, batch norm folding, etc.) to fit your models on tiny devices, we’d love to hear from you! Contact us at
Latency on NXP® NPU with less than 1% of accuracy drop
| Model Name | <p align="center"> Latency on NXP® NPU | <p align="center"> Memory Size Ratio [ fp32/int8 ] |
|---|---|---|
| Mobilenet V2 | <p align="center"> 6.98 ms <p> | <p align="center">x 3.6 |
| Resnet50 | <p align="center">19.07 ms | <p align="center">x 3.9 |
| Efficientnetlite b0 | <p align="center">7.54 ms | <p align="center">x 3.7 |
| YoloV3 | <p align="center">129 ms | <p align="center">x 3.75 <p> |
Known issue
Quantized Gemm is not runnable with the rtm format, we will need to change those into Matmul and Add operations
Thanks
All those benchmarks where made using NXP® eiQ<sup>TM</sup> Portal! Thanks a lot NXP®
Contact us
Feel free to reach out by email at hello at datakalab.com