End-to-end cloud platform for neural network compression and hardware synthesis via hls4ml. From dataset to bitstream, no HLS code required.
A unified platform covering every stage of the ML-to-FPGA pipeline.
Native support for MNIST, CIFAR-10, and custom CSV uploads. Integrated AI Dataset Agent recommends optimal preprocessing strategies.
Define Baseline, Student, and QKeras architectures via GUI. Import pre-trained models. AI Architect designs board-optimized topologies.
Pruning, Knowledge Distillation, QAT (TF-MOT & QKeras), Low-Rank Factorization, and fused KD+QAP pipelines.
Direct Keras-to-C++ conversion via hls4ml with control over precision, reuse factors, and implementation strategies.
Surrogate FPGA resource estimator (LUT, FF, DSP, BRAM) calibrated against Vivado/Vitis HLS synthesis results.
Build Agent bridges cloud dashboard with on-premise Xilinx tools. Auto-generated DMA/AXI-Stream wrappers for Zynq SoC.
Seven steps from raw data to synthesizable FPGA firmware.
Upload dataset, AI recommends preprocessing
Define model via GUI or AI Architect
Float32 training with K-Fold CV
Pruning, KD, QAT, SVD, or fused pipelines
Accuracy vs. Size vs. Params
Keras to C++ via hls4ml
DMA wrappers + Build Agent
| Technique | Engine | Description | Tier |
|---|---|---|---|
| Baseline Training | TensorFlow/Keras | Float32 reference training with K-Fold CV | Free |
| Pruning | TF-MOT | Unstructured & structured weight sparsity | Free |
| Knowledge Distillation | Custom | Teacher→Student transfer with temperature scaling | Developer |
| QAT (TF-MOT) | TF-MOT | Post-training quantization-aware fine-tuning | Developer |
| QAT / QAP (QKeras) | QKeras | Per-layer bit-width control | Developer |
| KD + QAP (Fused) | QKeras + Custom | Combined distillation and quantization | Developer |
| Low-Rank (SVD) | NumPy/Keras | Matrix decomposition of dense layers | Developer |
| Automated Pipeline | All | Multi-stage sequences with score-based ranking | Developer |