AI compute platform
A co-designed hardware and software stack. Two families of purpose-built edge AI processors carry the inference workload — letting a single appliance analyse many channels at a fraction of the power a GPU server would draw.
Physical AI semiconductors
Chips, M.2 modules, PCIe accelerator cards and complete edge AI systems for high-density video inference.
| Product | Perf. | Power |
|---|---|---|
| DX-M1 chip / M.2 | 25 TOPS | 1–5 W |
| DX-M1M chip / M.2 | 25 TOPS | 3 W |
| DX-H1 Quattro PCIe | 100 TOPS | 20 W |
| DX-H1 V-NPU + codec | 50 TOPS | 30 W |
| DX-AIPlayer edge box | 25 TOPS | 15–37 W |
| DX-M2 2 nm, edge LLM | 80 TOPS | ~5 W |
MLSoC for physical AI
One heterogeneous chip runs the entire multimodal pipeline — 50 TOPS under 10 W, no glue silicon or external accelerators.
| Product | Perf. | Power |
|---|---|---|
| Modalix SoM, Jetson-compat. | 50 TOPS | <10 W |
| Modalix HHHL PCIe Gen5 | 50 TOPS | <10 W |
| Modalix MLSoC chip-down | 50 TOPS | <10 W |
| Modalix DevKit platform | 50 TOPS | <10 W |
VisionX abstracts both accelerator families behind a single model runtime and rule engine. Scale from a compact DEEPX camera to a SiMa.ai Modalix appliance — without changing the operator interface, event schema or integration APIs.
Modalix advantage
Accelerates the entire LMM pipeline on one chip — sensor capture, preprocessing, vision encoder, LMM and app logic.
Custom models
Standard ONNX / PyTorch / TensorFlow converts to optimized .dxnn or .sima NPU binaries in hours.
Deployment enquiry
Book a live demo, request a dev kit or size a deployment with our engineering team.