Open-source Embedded Software • Edge AI • FPGA AI Acceleration

NetKit Labs is an independent research and development initiative focused on embedded software, Edge AI, digital signal processing, and FPGA acceleration for resource-constrained intelligent systems.

Our work spans bare-metal and RTOS firmware on ARM Cortex-M, embedded Linux, real-time DSP, TinyML inference runtimes, and FPGA-based neural network acceleration — practical open-source tools from microcontrollers to embedded Linux and programmable logic.

• Embedded Software Development (C/C++)
• ARM Cortex-M Firmware
• Bare-Metal & RTOS Systems
• Embedded Linux
• Digital Signal Processing (DSP)
• Edge AI / TinyML
• Neural Network Inference Runtimes
• FPGA AI Accelerators
• Hardware/Software Co-Design

Built and open-sourced NetKit, a custom C++ inference runtime for MCU and embedded Linux targets, with optimized INT8 inference kernels achieving up to 10× the performance of TensorFlow Lite for Microcontrollers (TFLM) reference implementations on selected operators. Float32 and INT8 paths are peer-benched across Arm MCU/MPU, Espressif, RISC-V, and host CPU backends (CMSIS-NN, ESP-NN, NMSIS-NN, XNNPACK).

Repository: NetKit-Labs/netkit

Highlights include:

• Portable C++ implementation
• INT8 quantized inference
• Deterministic memory management
• Resource-constrained deployment
• Optimized inference kernels
• Up to 10× faster than TensorFlow Lite for Microcontrollers (TFLM) reference kernels on selected operators

Open-sourced MemKit, an embedded systems container library written in C++ with native C and C++ APIs over shared implementations, designed specifically for resource-constrained MCU and MPU environments with predictable static, arena, and optional heap-backed storage.

Repository: NetKit-Labs/memkit

Features include:

• STL-inspired interfaces
• Predictable memory usage
• Native C and C++ APIs
• No dynamic allocation requirements during runtime
• Designed for real-time embedded applications

Designed and implemented an FPGA-based INT8 neural network accelerator in SystemVerilog featuring a streaming output-stationary systolic-array architecture for matrix multiplication, enabling quantized Vision Transformer inference on an AMD-Xilinx Zynq-7000 SoC.

Repository: NetKit-Labs/npukit

Highlights include:

• Output-stationary systolic-array architecture
• Streaming matrix multiplication engine
• Quantized neural network acceleration (INT8)
• AMD-Xilinx Zynq-7000 implementation
• Hardware/software co-design for embedded AI systems
• Host-scheduled tiny-ViT on MNIST with MCU-class DS-CNN peer benchmarks

Languages

• C
• C++
• SystemVerilog
• Python

Platforms

• ARM Cortex-M
• Embedded Linux
• AMD-Xilinx Zynq SoCs
• FPGA

Domains

• Embedded Systems
• DSP
• TinyML
• Edge AI
• Computer Vision
• Neural Network Acceleration

NetKit Labs exists to advance open-source Embedded AI by developing practical software infrastructure and hardware acceleration techniques that bridge embedded systems and modern machine learning.

The long-term vision is to provide a complete software-to-silicon ecosystem for efficient AI inference on resource-constrained devices.

James Lavrenz

Principal Embedded Software Engineer (Edge AI/TinyML) | Embedded AI | FPGA AI Acceleration

LinkedIn: https://www.linkedin.com/in/jameslavrenz