Hardware and Software Compatibility Matrix This document describes the hardware, software, and platform compatibility for SecAI_OS.
Last updated: 2026-05-14
Component
Value
Base image
Fedora Silverblue 44 (uBlue)
Image type
Immutable (rpm-ostree)
Builder
BlueBuild with cosign signing
Vendor
GPU Family
Backend
LLM Support
Diffusion Support
Status
NVIDIA
RTX 50-series
CUDA
Code path
Code path
Hardware-dependent
NVIDIA
RTX 40-series
CUDA
Code path
Code path
Hardware-dependent
NVIDIA
RTX 30-series
CUDA
Code path
Code path
Hardware-dependent
AMD
RDNA3
ROCm/HIP
Code path
Code path
Experimental
AMD
RDNA2
ROCm/HIP
Code path
Code path
Experimental
AMD
CDNA
ROCm/HIP
Code path
Code path
Experimental
Intel
Arc A-series
XPU/Vulkan
Code path
Code path
Experimental
Intel
Arc B-series
XPU/Vulkan
Code path
Code path
Experimental
Apple
M4/M3/M2/M1
Metal/MPS
Dev/sandbox
Dev/sandbox
No native appliance
Any
CPU only
AVX2/AVX-512
Yes
Yes (slow)
Supported fallback
“Hardware-dependent” becomes “Verified” only when a release publishes the
exact hardware, firmware, driver, runtime, workload, duration, and result.
Component
Value
LLM inference
llama.cpp (llama-server)
Model format
GGUF
Multi-GPU
Supported (tensor splitting across devices)
Resource
Minimum
Notes
RAM
16 GB
32 GB recommended for larger models
VRAM
8 GB
For GPU-accelerated inference
Storage
64 GB
SSD strongly recommended; more needed for multiple models
CPU
x86_64 with AVX2 or ARM64
AVX-512 preferred for CPU inference
Recommended Specs by Workload
Workload
RAM
VRAM
Storage
Notes
Small LLMs (7B parameters)
16 GB
8 GB
64 GB
Runs on most modern hardware
Medium LLMs (13B-30B parameters)
32 GB
16 GB
128 GB
RTX 4070+ or equivalent recommended
Large LLMs (65B+ parameters)
64 GB
24 GB+
256 GB
RTX 4090/5080 or multi-GPU setup
Image generation (diffusion)
32 GB
12 GB
128 GB
Dedicated VRAM for diffusion models
Multi-model serving
64 GB
24 GB+
512 GB
Multiple models loaded simultaneously
Platform
Format
Status
Notes
VirtualBox
Local OVA build
Evaluation
GPU passthrough requires compatible host
VMware (Workstation/ESXi)
Local OVA build
Evaluation
vGPU or passthrough for GPU acceleration
KVM/QEMU
Local QCOW2 build
Evaluation
VFIO passthrough for GPU acceleration
Proxmox
Local QCOW2 build
Evaluation
PCI passthrough for GPU acceleration
GPU passthrough is required for GPU-accelerated inference inside a VM.
CPU-only inference works in any VM without passthrough.
VM images are built locally so encrypted-boot credentials are user-specific;
generic QCOW2/OVA release artifacts are not published.
Allocate at least 16 GB RAM to the VM.
Nested virtualization is not required.
Secure Boot in VM requires the host to support UEFI boot for the guest.