xnor develops specialized edge AI solutions for efficient, low-power machine learning deployment. The company focuses on optimizing deep neural networks for resource-constrained hardware environments. This enables real-time inference capabilities directly on devices without relying on cloud connectivity.
Funding
$12M raised to dateRaised to date based on public sources. This may differ from the amount the company actually raised and is based only on what is publicly available on the internet.
MFounders
Product
Problem
Deploying deep learning models on edge devices such as IoT sensors, wearables, and autonomous robots is hampered by limited compute, memory, and power budgets. Conventional cloud‑based inference introduces latency, bandwidth costs, and data‑privacy concerns, preventing real‑time decision making at the source.
Solution
xnor delivers an edge‑AI platform that transforms high‑accuracy neural networks into ultra‑efficient models capable of running on resource‑constrained hardware. The workflow includes automated quantization, pruning, and architecture‑aware optimization that reduce model size and arithmetic intensity while preserving performance. A hardware‑aware compiler generates low‑level kernels tailored to microcontrollers, ASICs, and low‑power GPUs, producing a runtime library with a memory footprint measured in tens of kilobytes. The solution integrates with popular frameworks (TensorFlow, PyTorch, ONNX) and provides a cross‑platform SDK for seamless integration into embedded firmware. By enabling on‑device inference, xnor eliminates the need for continuous cloud connectivity, delivering sub‑millisecond latency and reducing power consumption for real‑time applications.
Target Audience
The primary customers are embedded system manufacturers, IoT device developers, and robotics firms that require on‑device AI capabilities, as well as AI engineers seeking to deploy low‑power inference across heterogeneous edge hardware.
Features
- Post‑training quantization to 8‑bit, 4‑bit, and binary formats with minimal accuracy loss
- Structured pruning and neural‑architecture search that automatically adapt models to target device constraints
- Hardware‑aware code generation for ARM Cortex‑M, RISC‑V, low‑power GPUs, and custom ASICs, producing optimized SIMD kernels
- Lightweight inference runtime (< 100 KB RAM) supporting CNNs, RNNs, and transformer blocks
- End‑to‑end conversion pipeline from TensorFlow, PyTorch, and ONNX models with one‑click deployment scripts
- Secure OTA update mechanism for model patches and firmware upgrades
- Integrated profiling tools for latency, memory, and energy consumption analysis on target hardware