Reefy uses one image across small mini-PCs, repurposed laptops, custom GPU workstations, and rack servers. Hardware support still depends on the Linux kernel, included firmware, and the workload you intend to run.
Baseline requirements
| Component | Minimum | Practical guidance |
|---|---|---|
| CPU | x86-64 | Modern Intel and AMD systems are the primary target |
| Memory | 4 GB | Use 16 GB or more for several apps or local models |
| Boot media | 8 GB USB | A reliable USB 3 drive improves boot and update speed |
| Network | Ethernet or supported Wi-Fi | Ethernet is recommended for first boot and servers |
| Data storage | Optional internal SSD or NVMe | Recommended for models, recordings, and persistent app data |
ARM systems are not supported by the current public image.
CPU and memory
AI agents that call remote model providers need much less compute than local inference. A small four-core mini-PC with 8 to 16 GB of RAM can run agents, development tools, monitoring, and home applications.
Local language models need enough memory for model weights, runtime overhead, and the desired context window. Size the machine for the model rather than for Reefy itself.
GPU and NPU support
Reefy supports NVIDIA GPUs, AMD GPUs, and Intel GPUs and NPUs. Vendor-specific host drivers are activated only when an installed app requests them. The request is best effort: an app still starts when its requested accelerator is not installed or cannot be activated, so the app should provide a CPU or other fallback when practical.
| Accelerator | App request | Reefy host support | App image supplies |
|---|---|---|---|
| NVIDIA GPU | gpu: true or nvidia.com/gpu=all |
Matched open kernel modules, GSP firmware, host tools, Vulkan/EGL libraries, and CDI | CUDA, TensorRT, PyTorch, Isaac Sim, or another application runtime |
| AMD GPU | amd.com/gpu=all |
Matched external amdgpu modules, firmware, amd-smi, and CDI |
ROCm, HIP, Mesa, media, or another application runtime |
| Intel GPU | intel.com/gpu=all |
Matched i915 and xe modules, GPU firmware, and CDI |
OpenVINO, Level Zero, media, or another application runtime |
| Intel NPU | intel.com/npu=all |
Matched intel_vpu module, NPU firmware, and CDI |
OpenVINO or another NPU runtime |
One app may request more than one provider. Different apps may also use different vendors on the same machine, including an NVIDIA and AMD GPU at the same time.
Accelerator compatibility is the intersection of three things: Reefy's host provider must support the device, the runtime inside the app must support it, and the workload must support that runtime. A GPU supported by the host driver is not automatically supported by every CUDA, ROCm, OpenVINO, or media app.
NVIDIA GPUs
The current provider uses NVIDIA's R595 production branch. Reefy downloads the provider built for the exact Reefy OS build and activates it without a separate installation step. Because Reefy uses NVIDIA's open kernel modules, the supported class is Turing and newer. This includes corresponding GeForce RTX, RTX professional, and data-center products listed by NVIDIA. Consult NVIDIA's kernel module compatibility guide and the R595 supported-product table when checking an exact model.
Available VRAM often matters more than peak gaming performance for local language models. If a model does not fit completely in VRAM, the runtime may use system memory or CPU execution with a substantial performance cost.
AMD GPUs and ROCm
Reefy supports ROCm applications on compatible AMD GPUs. The current host
provider is built from AMD GPU Driver 31.40 and includes the matching firmware,
CDI integration, and the amd-smi diagnostic. ROCm and HIP remain in the app
container, which keeps the host small and lets each app select its own runtime.
ROCm application support is narrower than the host driver's hardware reach and varies by ROCm release. Check AMD's current Radeon and Ryzen compatibility documentation and ROCm compatibility matrix for the runtime shipped by the app. These matrices cover supported Radeon, Ryzen AI, and Instinct families without implying that every AMD GPU supports ROCm compute.
Intel integrated GPU and NPU
Reefy activates Intel graphics and NPU kernel modules and firmware from one build-matched, on-demand provider. The provider publishes separate GPU and NPU CDI resources only for usable devices. Application support still depends on the runtime. The Reefy video benchmark includes OpenVINO GPU and NPU paths so performance and detection parity can be measured rather than assumed.
Intel GPU support spans the i915 and Xe driver families included in the build-matched provider. Intel publishes an exact Xe driver GPU table. For NPU applications, consult the verified platforms in the Intel Linux NPU driver releases and the application's runtime requirements, such as the OpenVINO system requirements. Measured OpenVINO results for Reefy's tested Intel GPU and NPU configuration are available in the AI video inference benchmark.
Validated by Reefy
Vendor-supported hardware has not necessarily been tested by Reefy. The following configurations are part of Reefy's real-hardware E2E coverage for accelerator-provider releases:
| Configuration | Reefy E2E coverage |
|---|---|
| NVIDIA GeForce RTX 5060 Ti | Provider activation, nvidia-smi, CDI delivery, CUDA compute, and NVIDIA application startup |
| AMD Radeon RX 9060 XT | Provider activation, amd-smi, CDI delivery, ROCm/HIP compute, and AMD application startup |
| Intel Lunar Lake GPU and NPU | Provider activation, separate GPU and NPU CDI delivery, OpenVINO compute on both devices, and Intel application startup |
| NVIDIA and AMD in one system | Both providers active concurrently and simultaneous compute workloads without CDI conflicts |
The hardware E2E suite also exercises provider download and caching, reboot activation, app reconciliation, and detection of provider or firmware mismatches. A device absent from this table may work, but Reefy has not run the same release-gated tests on that exact configuration. Laptop firmware, hybrid graphics, BIOS settings, PCIe topology, and the runtime packaged by an app can still affect compatibility.
The GPU and NPU provider architecture explains driver delivery, compatibility, activation, and fallback behavior.
Storage
Running Reefy from USB leaves internal drives available for encrypted data. Adopted storage uses LUKS2 encryption, LVM thin provisioning, and XFS. Fast NVMe storage helps model loading, video recording, development environments, and backup snapshots.
The USB drive is operationally important when it contains key material. Keep a recovery plan and do not treat an unknown low-quality flash drive as permanent server media.
Firmware settings
- Use UEFI boot mode.
- Disable Secure Boot for the current public image.
- Enable hardware virtualization when applications need KVM.
- Enable the integrated GPU if you plan to use it alongside a discrete GPU.
- Configure automatic power recovery if the device should return after an outage.
What "supported" means
The public image targets a broad x86-64 hardware range, but it cannot promise that every peripheral in every PC has a working upstream driver. A machine is confirmed only after its storage, network, sensors, accelerators, suspend behavior, and update path have been exercised.