Which GPU do I need?

Everything here is about the Moonlight PC, the PC that runs Moonlight, not the PC you play on.

Color runs no AI model and needs no graphics card. AI and Console run a detection model. A graphics card makes that faster, and its make and generation decide which Backend you can use on the Detection page.

#Pick by graphics card

The minimum cards run the lighter, faster models such as the 320 ones. They keep detection fast, but they find targets less reliably than the larger models stronger hardware can run. Better hardware buys detection speed and room for bigger models.

Roughly 100 detections per second is enough to work with. Ideally detection FPS is at or above your capture FPS. The Live Preview shows both.

#TensorRT

TensorRT runs only on NVIDIA GTX 16 series, RTX 20 series and newer, laptop versions included. It needs NVIDIA driver 580 or newer. You do not need to install CUDA; Moonlight ships what it needs.

The first Start Moonlight with each model builds an engine for your card. It can show no picture for several minutes. Later starts are fast. Changing the model or Input Size builds again. Keep TRT Optimization at Level 3 unless you want to test others, and keep about 2 GB free on Moonlight's drive.

On a GTX 10 series or older card, AMD or Intel, use ONNX. Moonlight may still list TensorRT, but it will not build on those cards.

#No graphics card

Choose OpenVINO as the Backend and CPU or AUTO as the OpenVINO Device, with a 320 model. ONNX does not run without a graphics card.

These times are estimates.

#AMD and Intel

AMD cards use ONNX, which runs on DirectML. TensorRT is NVIDIA only, so not seeing it work on AMD is expected.

Intel Arc and Intel integrated graphics can use ONNX or OpenVINO with OpenVINO Device set to GPU. On Intel Core Ultra PCs, NPU appears as a device once Moonlight has checked that it works.

Desktop Ryzen 7000 and 9000 processors without a G in the name have a very small graphics chip. Treat them as no graphics card.

#Laptops

A laptop works as the Moonlight PC. The same rules apply by graphics chip: an RTX or GTX 16 laptop uses TensorRT, a GTX 10 or MX laptop uses ONNX. Laptop graphics are slower than the desktop card with the same name, so start one model size smaller.

• Keep it plugged in. On battery the GPU and CPU are slowed down.
• Make Windows run Moonlight on the fast graphics chip: see Laptops and PCs with two graphics chips.
• Plug a capture card into a 10 Gbps USB port, or it is limited to about 120 FPS.

#VRAM

Moonlight sets no VRAM minimum. The detection models are small, so a 2 GB card such as the GTX 1050 runs them. How fast the card is, and whether it can use TensorRT, matters more than how much VRAM it has.

#Model size

Smaller models and input sizes are faster; larger ones find small and distant targets more reliably. 320 is much faster than 640. In model names, n and p are the fastest sizes; s, m, l and x are more accurate and need more from the GPU.

Input Size can only be changed on models with "dynamic" in the name. Stop and Start after changing the model or backend.

#Processor and memory

Every product needs a CPU with AVX2 support. Every Ryzen and every Intel Core from 4th gen on has it; some Pentium and Celeron chips do not. Without it Moonlight does not run.

If detection is slow on hardware that should be enough, see Low detection FPS.