Speakers
Description
ProtoSDK is a low-code software development kit (SDK) aimed at simplifying the deployment of deep learning (DL) models on FPGA-based devices for space-related applications. By removing the necessity for specialized programming knowledge, ProtoSDK enhances accessibility and facilitates seamless workflows from selecting a model to deploying it on an FPGA.
The SDK enables an end user to select from predefined model architectures, configure relevant hyperparameters, and start the training process. Additionally, the level of quantization is selected. After the training, a benchmark compared to the baseline unquantized model is compared to the quantized output to determine its effectiveness. Such a model is then passed onto the streamlining and synthesis engine to carry out the necessary transformations to generate and synthesize the necessary IP cores that can be integrated into FPGA design. Depending on the target FPGA platform, either a complete bitstream file can be generated or standalone IP cores that require manual stitching into the existing FPGA design. Such deployments are further validated and benchmarked on hardware devices to ensure reliability, performance, and power usage in resource-constrained environments such as onboard data processing on satellite platforms, where power consumption and memory are limited.
ProtoSDK enables rapid prototyping and faster time to market with various state-of-the-art machine learning models, accelerated with FPGAs. This enables better data processing capabilities and lowers downlink costs.
Affiliation of author(s)
Protostar Labs
| Track | Artificial Intelligence/Machine Learning |
|---|
