Hardware development platform · FPGA + AI
ZU-NEXUS-AI
Advanced AI Vision for Embedded Systems
The strategic nexus between multiprocessor FPGA technology and neural acceleration: a flexible, high-performance hardware development platform combining the Zynq UltraScale+ SoC with a dedicated neural accelerator.
Three hardware components, one integrated platform
ZU-NEXUS-AI integrates three elements engineered by KED to work together.
01 · Compute module
SOM Zynq UltraScale+
- ZU1EG / ZU2EG / ZU3EG SoC
- 4× ARM Cortex-A53 + 2× Cortex-R5 lockstep
- LPDDR4 32-bit, eMMC + QSPI Flash
- 4× MGT (PCIe / USB3 / DisplayPort)
- SODIMM 200-pin form factor, 67.6 × 30 mm
02 · Carrier board
ZU-NEXUS-AI Board
- USB-C dual-role (USB 3.0) + USB-C JTAG/console
- DisplayPort · Gigabit Ethernet · SD Card
- M.2 Key-M PCIe ×2 for AI accelerators
- MIPI CSI-2 cameras (22-pin ZIF)
- Standard SYZYGY expansion · 100 × 100 mm
03 · Neural accelerator
DeepX NPU module
- DX-M1 8-core NPU, 25 TOPS INT8
- PCIe Gen2 ×2 interface
- DX-M1 in M.2 2280 Key-M
- DX-M1M in M.2 2242 Key-M
- TDP < 5 W · ONNX / TFLite / PyTorch SDK
The processing pipeline
Optimal workload distribution across FPGA, CPU and neural accelerator.
Sensor
MIPI / Camera Link, radar, ultrasound
Pre-processing
Resize, conversion, DSP filters
DDR buffer
Double / triple buffering
Inference
INT8 neural networks
Post-processing
Decode, fusion, tracking
Deterministic pre-processing
The FPGA handles acquisition and conversion at fixed, clock-cycle-accurate latency.
Offloaded CPU
Heavy image operations do not burden the ARM cores, which remain free for application logic.
Dedicated NPU
The accelerator runs almost exclusively pure inference, with high and constant throughput.
Multi-sensor fusion
Post-processing fuses camera, radar and other sensors into a single coherent stream.
Board interfaces and specifications
All the SOM high-speed resources exposed on standard connectors.
- SoC
- Zynq UltraScale+ ZU1EG / ZU2EG / ZU3EG
- APU
- 4× ARM Cortex-A53
- RPU
- 2× ARM Cortex-R5 (lockstep)
- RAM
- LPDDR4 32-bit
- Storage
- eMMC + QSPI Flash + SD Card
- NPU
- DeepX DX-M1 · 25 TOPS INT8
- NPU interface
- PCIe Gen2 ×2 (M.2 Key-M)
- Video out
- DisplayPort 1.x — 4K@30 / 1080p@60
- Power
- 5–12 V DC · 8–12 W typical
- Dimensions
- 100 × 100 mm
- Expansion up to 3 MIPI CSI-2 cameras
- USB-C dual-role USB 3.0 SuperSpeed
- USB-C JTAG + dual UART console (FTDI)
- Gigabit Ethernet (RGMII)
- M.2 Key-M PCIe ×2 for AI modules
- SYZYGY and PMOD expansion connectors
- Compact SWaP-optimized form factor
One platform, five high-value sectors
The same architecture adapts to domains with very different requirements.
Medical
Diagnostic imaging: ultrasound, endoscopy, diagnosis support
Railway
Camera + radar sensor fusion, ATO, infrastructure inspection
Industrial
Quality control, machine vision, real-time robot guidance
Drones / UAV
On-board detection and tracking, ISR, mapping with EO/IR payloads
Space
Vision payloads for satellites and CubeSats, on-board processing
The KED added value
- Complete Linux Board Support Package for the Zynq UltraScale+ platform
- Custom embedded distributions: Yocto, Buildroot, Debian
- Custom operating systems and RTOS configurations for real-time requirements
- Driver integration, hardware bring-up and boot optimization
- Custom HDL IP for video acquisition and pre-processing
- Signal and image processing pipelines in programmable logic
- AI accelerator integration into the FPGA data-path
Documentation
- Hardware Architecture Document — Rev. 1.0PDF · coming soon
- ZU-NEXUS-AI brochurePDF
- General terms and conditions of salePDF · coming soon
- Technical wiki: TRM, reference designs, Buildroot, GStreamerWIKI · coming soon