Written in Rust.
Built for Radeon.
Made for Moonlight.

Measured, not claimed

Measured picture age under GPU load

The October 12 native-HDR tests measured 14.34 ms average picture age for HEVC and 13.65 ms for AV1 at 1968×2184 and 120 fps. A timestamped picture passes through capture, encoding and independent hardware decoding on the same Radeon PC.

Rubylight 2.2.9 — October 12, 2026. The refresh passed 81 encoder probes, encoded and scored 148 AMF quality cases (32 defaults plus 116 settings variants), and passed 36 short loopback streams and 12 static-keyframe checks, alongside CPU, GPU, input and PyroWave component checks. Current measurements and scope

The AMF property sweeps passed 1,018/1,030 probes. Twelve 4K SDR cases with preanalysis enabled failed completion or frame-accounting checks; these probes do not measure full-stream latency. Settings results and failures

Both 4K120 HEVC runs and both 5120×1440/240 FPS saturation runs failed the stream gates. The audio/load/outage batch passed 6/11 full-stream checks, while 11/11 retained continuous audio. Those earlier stream batches used physical SDR capture with converted HDR output, separately from the native-HDR measurements.

Both 120-second audio streams passed, with 24/24 continuous audio windows. All four loss controls passed; all four injected-outage runs failed the normal gap checks. The short soak passed 64/67 cases, with three GPU-load freshness failures. Reconnect, switching, helper recovery and host working set, handle and thread growth checks passed. A 30-reconnect follow-up reproduced the rising GPU process counter as a reporting artifact: adapter-wide memory returned to baseline, with no sustained VRAM growth observed. Soak results and scope

The real Radeon 780M Wi-Fi receiver passed 20/20 stream checks: 12 HDR-output conversion runs and eight control/recovery runs. 19/20 calibrated timing records passed source validation; one had mismatched barcode timestamps and remains unavailable. Timing ends at frame assembly/enqueue, before decoder queueing, decoding or display. Current Wi-Fi results

Official Moonlight 6.2.0 completed two 165-second 1080p60 HEVC SDR sessions requesting 80 Mbps on the Radeon 780M laptop over Wi-Fi, reporting 59.99 FPS rendering in both. These component statistics do not measure full picture age. Both wrappers failed their immediate exit check; later checks confirmed natural exit with settings and pairings preserved. Stock Moonlight results

The native-HDR follow-up passed 28/28 DDX/WGC stream and timing checks and 4/4 HDR pixel checks. The separate SDR VRR capture sweep has 20 valid requested-mode phases; 12 phases remain unavailable because Windows selected different refresh rates. The historical private HDR/RTSS fixture source is unavailable.

Rubylight 2.2.9 · October 12 Vibepollo 2.0 · October 8

1968×2184 HEVC HDR, 120 fps

14.3 msp95 16.0 Rubylight 2.2.9, October 12: average 14.34 ms, p95 16.03 ms, p99 18.15 ms, host latency 3.35 ms

27.0 msp95 37.6 Vibepollo 2.0: average 26.96 ms, p95 37.62 ms, p99 39.25 ms, host latency 8.3 ms

1968×2184 AV1 HDR, 120 fps

13.6 msp95 18.0 Rubylight 2.2.9, October 12: average 13.65 ms, p95 18.00 ms, p99 18.95 ms, host latency 5.25 ms

24.5 msp95 36.8 Vibepollo 2.0: average 24.53 ms, p95 36.77 ms, p99 38.05 ms, host latency 7.4 ms

1920×1080 HEVC HDR, 60 fps

10.6 msp95 18.3 Rubylight 2.2.9, October 12: average 10.64 ms, p95 18.27 ms, p99 18.58 ms, host latency 2.24 ms

20.6 msp95 22.6 Vibepollo 2.0 (one run): average 20.55 ms, p95 22.55 ms, p99 22.66 ms, host latency 5.7 ms

010203040

Current DDX results average two runs per codec and include local hardware decoding, excluding remote network and display time. The virtual HDR source remains at 120 Hz. The physical desktop changed from 2560×1440/120 Hz on October 8 to 5120×1440/240 Hz on October 12. Different dates and desktop conditions prevent a matched improvement claim. Current native runs and scope

Historical matched comparison: October 8, 2026

Rubylight rc.24 and the pinned Vibepollo 2.0 build on the same GPU, alternating in one batch, two runs per cell. Identical settings: native AMF at ultra-low latency, Desktop Duplication capture, a virtual HDR display. The picture is decoded in hardware by an independent client on the same PC, which isolates the host from the network so the two can be compared precisely; the Wi-Fi results below add a real network. Beside the game Rubylight also delivered more fresh pictures: 117–120 a second against 103–118. Runs and settings

Over Wi-Fi

The October 12 Rubylight 2.2.9 hardware-receiver matrix passed 20/20 ordinary stream checks, with 19 valid calibrated timing records. It measures arrival at assembly/enqueue from a physical SDR source, including HDR-output conversion. The 1.177 ms clock-fit diagnostic is not a proven error bound. The following October 8 measurements retain their original method and date.

A client on 5 GHz Wi-Fi, decoding in hardware, against the host on Ethernet. "Received" is when the client starts decoding a fully received picture.

Rubylight rc.24, October 8, 2026, one 30-second run per row. The client's clock was synchronised to the host with 200 ms pings (error about ±0.6 ms). Every picture arrived and decoded. Method and clock sync

With Moonlight on the same link, a 1440p120 AV1 stream held 119.7 fps with no loss, and with the host dropping 20 ms of video every 1.5 s, all 59 lost frames were recovered without a keyframe at 117.7 fps.

Earlier results

• Radeon compute, one setting changed: average picture age beside a game-like load fell from 41.0 to 33.5 ms and the 95th percentile from 54.4 to 42.3 ms (1080p60 HEVC HDR). Details
• PyroWave beside a GPU-heavy game: 0.55 ms per 1080p HDR 4:4:4 frame instead of 5.7 ms, now that its colour conversion runs on the compute queue. Details
• A saturated encoder: at 5120×1440 HEVC 240 fps a game frame reaches the network in 11.1 ms instead of 42.7 ms. Details
• Input under CPU load: a mouse move behind a controller update took at worst 2.0 ms instead of 287 ms. Details
• H.264 picture quality: adaptive quantization off by default raised VMAF by 1.0 and 9.4 points at 1440p120 and 20 Mbps, at the same encode time. Details

Works with

• Every Moonlight and Sunshine-compatible client, on Windows, macOS, Linux, Android, iOS, TVs, consoles and handhelds.
• AMD Radeon GPUs, desktop and mobile, as the host.
• Locked Windows, so you can sign in from the couch, headless machines and VMs with no monitor, and Kubernetes pods with GPUs.

All performance results and what each number means

Install. Pair Moonlight. Play.

Three steps to your first stream

1. Install

   Run the installer from the release. Installing over Apollo, Vibepollo or Butterpollo keeps your settings, paired devices, app library and covers.

2. Open the console

   Go to https://localhost:47990 on the host PC and create your local account.

3. Pair and play

   Add your PC in Moonlight or any Sunshine-compatible client, enter its PIN under Devices, then launch Desktop. Start at 1080p60, then raise resolution, frame rate and HDR.

First-stream walkthrough, portable setup and migration

What users say

From people running it on Radeon

Quotes from Reddit, lightly edited for typos. Read more of what users say

Documentation

Find what you need