Defines the hardware support matrix, software compatibility, support windows, and deprecation policy for SecAI OS.
Last updated: 2026-03-14
| Vendor | Generation | Models | Backend | Support Level | Notes |
|---|---|---|---|---|---|
| NVIDIA | Ada Lovelace | RTX 4090, 4080, 4070 Ti/Super | CUDA 12.x | Full | Primary development target |
| NVIDIA | Blackwell | RTX 5090, 5080 | CUDA 12.x | Full | Tested on RTX 5080 |
| NVIDIA | Ampere | RTX 3090, 3080, 3070 Ti | CUDA 12.x | Full | |
| AMD | RDNA 3 | RX 7900 XTX/XT, 7800 XT, 7700 XT | ROCm 6.x (HIP) | Full | |
| AMD | CDNA 3 | MI300X | ROCm 6.x (HIP) | Full | Data center GPU |
| Intel | Alchemist | Arc A770, A750, A580 | XPU (oneAPI) | Supported | Via Vulkan for LLM |
| Intel | Battlemage | Arc B-series | XPU (oneAPI) | Supported | Via Vulkan for LLM |
| Apple | M4 | M4, M4 Pro, M4 Max, M4 Ultra | Metal / MPS | Full | Via llama.cpp Metal |
| Apple | M3 | M3, M3 Pro, M3 Max, M3 Ultra | Metal / MPS | Full | |
| Apple | M2 | M2, M2 Pro, M2 Max, M2 Ultra | Metal / MPS | Full | |
| Apple | M1 | M1, M1 Pro, M1 Max, M1 Ultra | Metal / MPS | Full | |
| CPU | x86_64 | AVX2 or AVX-512 required | CPU | Supported | Slow but functional |
| CPU | ARM64 | NEON required | CPU | Supported | Slow but functional |
Support levels:
- Full: Actively tested, optimized, documented. Issues prioritized.
- Supported: Tested, functional. Issues accepted but lower priority.
- Community: Not actively tested. Community contributions welcome.
| GPU Vendor | Minimum Driver | Recommended Driver | Notes |
|---|---|---|---|
| NVIDIA | 535.x | 550.x or later | CUDA 12.2+ required |
| AMD | ROCm 5.7 | ROCm 6.0+ | HIP runtime |
| Intel | Level Zero 1.3 | Latest stable | oneAPI 2024.0+ |
| Apple | macOS 14.0 | macOS 15.0+ | Metal 3 for best performance |
| Resource | Minimum | Recommended | Notes |
|---|---|---|---|
| CPU | 4 cores, x86_64 or ARM64 | 8+ cores | AVX2 required for x86_64 |
| RAM | 16 GB | 32 GB+ | Swap is disabled by design |
| Storage | 64 GB SSD | 256 GB+ NVMe | For OS + models + vault |
| GPU VRAM | 8 GB | 16 GB+ | Determines max model size |
| Network | Optional | Ethernet | Only needed for initial setup / airlock |
| TPM | Optional | TPM 2.0 | Required for measured boot and key sealing |
| UEFI | Required | Secure Boot capable | Secure Boot optional but recommended |
| Install Path | Support Level | Security Features | Notes |
|---|---|---|---|
| Bare metal (ISO) | Production | Full: TPM2, Secure Boot, hardware isolation, fs-verity | Recommended for sensitive workloads |
| Bare metal (rebase) | Production | Full | For existing Fedora Silverblue operators |
| VM (OVA/QCOW2) | Evaluation | Limited: no TPM2 sealing, host visibility of VM memory | Not for sensitive model data |
| VM (manual) | Community | Limited | Self-configured |
| Container (dev) | Development | Minimal: no systemd hardening, no firewall, no vault | Service development only |
Bare-metal installations using the stable release channel with digest-pinned images are the intended production path. This means:
- Security patches within 72 hours of disclosure
- Automated rollback via Greenboot
- Supply-chain verification (cosign + SLSA3 provenance)
- Documented recovery procedures
VM installations are supported for evaluation and development only. Known limitations:
- No TPM2 vault key sealing (secrets held in VM memory visible to host)
- No Secure Boot chain verification (depends on hypervisor configuration)
- Reduced GPU performance (passthrough required for full acceleration)
- Host hypervisor has full visibility of guest memory
Note: The ISO install path is newly added and should be considered production-ready once exercised in at least one stable release cycle. Until then, the bare-metal rebase path remains the most validated production install method.
| Component | Version | Support Window | Notes |
|---|---|---|---|
| Fedora Silverblue | 42 | Until Fedora 42 EOL (~13 months from release) | Current base |
| uBlue framework | Latest | Follows Fedora lifecycle | Image build framework |
| rpm-ostree | 2024.x+ | Follows Fedora lifecycle | Immutable OS layer |
| Component | Version | Pinned | Notes |
|---|---|---|---|
| Go (services) | 1.25 | Yes (go.mod) | 9 Go services |
| Python | 3.12 | Yes (Fedora 42 default) | 6 Python services |
| llama.cpp | Latest stable | Via build | LLM inference engine |
| Flask | 3.x | Via pip | Web UI framework |
| cosign | 2.4.x | Via release workflow | Image/artifact signing |
| Syft | Latest stable | Via CI | SBOM generation |
| Backend | Model Format | Supported Versions | Notes |
|---|---|---|---|
| llama.cpp | GGUF | v3 header format | Primary inference engine |
| Diffusers | Safetensors | HuggingFace format | Image/video generation |
Unsupported formats: Pickle (.pt, .pkl, .bin) — rejected by quarantine stage 2 (Format Gate) due to arbitrary code execution risk.
| Release Type | Active Support | Security Patches | Notes |
|---|---|---|---|
| Current stable | Full | Full | Latest vX.Y.Z tag |
| Previous minor | Security only | 6 months after next minor | e.g., v1.2.x after v1.3.0 ships |
| Previous major | None | 12 months after next major | e.g., v1.x after v2.0.0 ships |
- Announce: Feature marked as deprecated in release notes and documentation
- Warn: Deprecated feature emits log warnings when used (minimum 1 minor release)
- Remove: Feature removed in next major version
Minimum deprecation window: 2 minor releases or 6 months, whichever is longer.
| API Change | Notice Period | Backward Compatible |
|---|---|---|
| New endpoint | Immediate | Yes (additive) |
| New optional field | Immediate | Yes (additive) |
| Remove endpoint | 2 minor releases | No (removed in next major) |
| Change field semantics | 1 minor release | Aliased during transition |
| Policy schema change | 1 major release | Migration tool provided |
| Schema Version | Compatible Releases | Migration |
|---|---|---|
| v1 (current) | v1.x | N/A |
| v2 (future) | v2.x | Migration tool + docs |
Policy files (policy.yaml, agent.yaml) are validated at startup. Invalid policy files cause fail-closed behavior (all requests denied) rather than silent degradation.
| From → To | Method | Automatic | Data Migration |
|---|---|---|---|
| Patch → Patch | rpm-ostree upgrade |
Yes | None |
| Minor → Minor | rpm-ostree upgrade |
Yes | None (new defaults) |
| Major → Major | rpm-ostree upgrade + migration |
Semi-auto | Per release notes |
| Rollback (any) | rpm-ostree rollback |
Yes | N/A (previous state) |
When a release reaches end-of-life:
- No further security patches
- No bug fixes
- CI may stop testing against that version
- Documentation may reference newer versions only
- Users should upgrade to a supported release
Notification: EOL announcements are made at least 30 days in advance via:
- GitHub Release notes
- README update
- Security advisory (if security-relevant)
- SecAI OS image and all bundled services
- Policy configuration within documented schema
- GPU acceleration for supported hardware (see matrix above)
- Upgrade/rollback via rpm-ostree
- Supply chain verification via cosign/SBOM
- Incident response automation
- Audit logging and integrity monitoring
- Custom kernel modules or drivers not in the base image
- Third-party services not bundled with SecAI OS
- Model training (inference only)
- Multi-node/cluster deployments (single-node appliance only)
- Cloud provider integrations
- Custom OCI image modifications (fork the recipe instead)
- Performance tuning beyond documented configuration
- Recovery from emergency panic level 3 (full wipe is unrecoverable by design)
- Non-standard GPU configurations
- Exotic hardware (FPGA, custom accelerators)
- Running in containers (not as a host OS)
- Non-x86_64/ARM64 architectures