One platform, two paths to hardware

Around 80% of our stack — applications, API, runtime — is hardware-independent. Beneath it sits an interchangeable hardware layer: our own Vexon-1 NPU or available edge hardware.

What is the NeuroVexon platform?

A platform is more than a chip. At NeuroVexon it means the complete software stack that makes AI applications run — from the finished application through the programming interface (API) to the runtime that actually executes the AI model.

What makes it special: this stack is decoupled from the hardware. The same software runs on our own Vexon-1 NPU just as it does on available edge hardware. The customer decides how deep to go into the sovereign hardware — without having to rewrite their application.

Sovereign Edge-AI — live in action

People, vehicles and surfaces detected in real time — faces anonymized on the spot (GDPR), entirely local on your own edge hardware. No cloud.

The shared stack

The same software, independent of the underlying hardware.

ApplicationsLLM chat · Sensing · Vision · custom apps
API layerREST API · SDK
RuntimeModel execution · memory management
Driver & LinuxOwn Linux image · C++ driver · DMA
Vexon-1 NPUOwn IP · RISC-V firmware
HardwareFPGA today · ASIC in preparation

One continuous stack — from the application down to our own hardware.

Development status — honestly

What runs in production today, what is proven in hardware, and what we are working on right now.

Applications, API, runtime
In production
In use on standard hardware: assistant, vision inference, REST API and SDK.
Own Linux image + C++ driver
Verified in hardware
Own PetaLinux image with our own driver stack; DMA data path bit-exact on two FPGA platforms.
NPU firmware (RISC-V sequencer)
Verified in hardware
Complete compute flow from firmware, no RTL change. The data schedule shipped with the core is already optimised: factor 1.3 over a simple command sequence, bit-identical result.
Compiler bridge (model → NPU)
In progress
Automatic mapping of models onto the NPU — toolchain decision pending.
Complete network end to end on the NPU
In progress
Language-model tokens shown with the NPU in the data path; full networks and sensor benchmarks to follow.

An interchangeable hardware layer

SOVEREIGN IP

Vexon-1 — the sovereign depth

Our own edge NPU with a RISC-V control core and the five core mechanisms, three of them registered as utility models — fully proprietary IP with a European supply chain.

  • In-house hardware driver (C++) → Vexon-1 RTL kernels
  • Today: verified in hardware on two FPGA platforms (AMD Kria KV260, Zynq UltraScale+ ZU15EG)
  • Goal: ASIC manufacturing with a European foundry partner
AVAILABLE TODAY

Industry-standard edge hardware

The same NeuroVexon stack also runs on available, industry-standard edge hardware. This already delivers a ready-made AI box (appliance) today — turnkey, vendor-independent and productive right away.

  • Same API, same runtime
  • Available as a turnkey AI box / appliance — today
  • Vendor-independent, no lock-in

The same stack, the same API — the customer chooses the hardware depth.

Discuss platform integration

Whether your own NPU or standard edge — we'll show you how the NeuroVexon stack fits your project.