Core Architecture
OmniStack operates as a dynamic overlay, hooking into the Linux kernel via eBPF and custom kmods. It adapts to the host distribution without requiring a custom kernel recompile per distro.
The stack at a glance
Hardware Abstraction & Silicon Translation
Instead of native drivers fighting for resources, OmniStack acts as a traffic controller for heterogeneous compute.
Intercepts matrix math and routes it to available GPU cores dynamically.
Edge CPU and Intel NPU optimization via OneAPI hooks.
Translates standard ML primitives into AMD-specific instructions.
Holds quantum states and offloads to IBM Qiskit, Google Cirq, or a local simulator.
Adaptive OS Layer (Distro Agnostic)
Ships as user-space daemons and kernel headers tailored per distribution family.
RHEL, SUSE SLES, Ubuntu Pro. Stability, NUMA-awareness, high availability.
Fedora. Testing ground for new scheduler algorithms.
VMware ESXi via VIBs for passthrough AI/Quantum acceleration to VMs.
Orchestration & Data Center
Multi-tenant, multi-rack, multi-region — auto-tuned to the hardware it lands on.
Auto-discovery agents spin up Kubernetes clusters tuned to the silicon they land on.
Hyperconverged storage and compute pooling via Ceph and NVMe-oF for AI datasets.
Networking Fabric · the nerve center
| Component | Integration strategy |
|---|---|
| SONiC | BGP EVPN control-plane integration to prioritize AI tensor-flow traffic across switches. |
| Arcos OS | Deep routing analytics for low-latency, lossless RoCEv2 transfer. |
| HPC Leaf & Spine | Zero-trust non-blocking topologies using RDMA over Converged Ethernet, tuned for Supermicro AI chassis. |
Toolkit distributions
OmniStack Enterprise
Data Center & HPC- Deployment
- Helm charts for K8s · RPM/DEB packages for bare-metal
- Scale
- Multi-node distributed AI training and hybrid-quantum algorithms
- Hardware
- Supermicro GPU/NPU multi-rack clusters · Equinix data centers
- Features
- VMware vSphere integration, live workload migration, automated failover, SLA-driven allocation
OmniStack Lite
Student & Developer Community- Deployment
- Docker Desktop extensions · Snap / Flatpak · Yocto BitBake images for Pi
- Scale
- Single-node optimization and simulation
- Hardware
- Consumer laptops · integrated GPUs · Google Coral and similar TPUs
- Features
- Simulated QPU · unified API wrapper: write once, compare CUDA vs. OpenVINO
Build & deployment strategy
Yocto Project (BitBake) recipes
Custom BitBake layers meta-omnistack-ai and meta-omnistack-qpu strip the kernel to essentials, compiling only the silicon drivers needed for edge TPU/NPU devices.
eBPF traffic-control optimization
eBPF programs intercept memory allocation requests. A matrix-multiply allocation is seamlessly redirected to GPU/NPU VRAM.
Kubernetes Device Plugins
Custom plugins teach K8s to see, request and limit CUDA cores, OpenVINO threads, and quantum qubits — not just CPU and memory.
Network OS overlay packaging
Networking daemons compiled as containerized apps that run natively inside SONiC and Arcos OS switches, managing Leaf/Spine congestion for AI telemetry.