Labs · Encoding
AV1 vs HEVC vs H.264 — what cloud gaming's codec migration actually buys you
The biggest invisible change in cloud gaming over 2024–2026 was the migration to AV1. We measured what it bought you — and where the marketing exaggerated.
The three codecs, in one paragraph each
H.264 (AVC) was the workhorse of streaming video and live game streaming for fifteen years. Every device on Earth decodes it. Its bitrate-to-quality ratio at 1080p+ is the worst of the three, but the decoder is the smallest and the fastest.
H.265 (HEVC) shipped widely starting 2015. ~35% bitrate savings at the same perceptual quality. License costs limited its web adoption, but Apple and the streaming-TV ecosystem embraced it.
AV1 is the Alliance for Open Media's royalty-free successor. ~20% better than HEVC at the same quality, with a heavier encoder. Hardware decoders are now standard on iPhone 15+, most 2023+ Android flagships, and the RTX 40-series GPUs already in NVIDIA's data centres.
Quality / bitrate / latency triangle
| Codec | Bitrate (1080p60 iso-quality) | VMAF | Encode | Decode | End-to-end |
|---|---|---|---|---|---|
| H.264 / AVC | 20 Mbps | 81.4 | 5.6 ms | 3.1 ms | 30.7 ms |
| H.265 / HEVC | 12 Mbps | 87.2 | 6.2 ms | 3.4 ms | 31.6 ms |
| AV1 | 9 Mbps | 89.1 | 7.4 ms | 4.0 ms | 33.4 ms |
Measured on a benchmark Cyberpunk 2077 scene at 1080p60, server-side encoder presets matched to each service's production configuration. VMAF computed against a lossless ground-truth render.
AV1 wins on quality at half the bitrate of H.264. Same scene, same perceived quality, 9 Mbps vs 20 Mbps. That's a real benefit if your connection is unstable or capped — there's more bitrate headroom for the encoder to react to packet loss.
AV1 costs about 3 ms on the latency budget. Encoder takes ~1.8 ms longer than H.264 and the decoder ~0.9 ms longer. End-to-end that's ~2.7 ms of extra input lag versus H.264 at the same scene — small but measurable, and meaningful when stacked against a 60 Hz frame budget of 16.7 ms.
HEVC is the practical sweet spot. Within 2 ms of H.264 on latency, within 2 VMAF points of AV1 on quality, decodes on essentially every device made in the last decade. The reason most services landed here first is that there isn't a strong reason to leave.
Where each service stands
- GeForce NowNow: AV1 (Ultimate), HEVC (Performance), H.264 (Free)Direction: Already migrated. AV1-by-default for all tiers expected late 2026.
- Xbox Cloud GamingNow: HEVC primary, AV1 in pilot regionsDirection: AV1 rolling out alongside the 1440p pilot through 2026.
- PlayStation Plus PremiumNow: HEVCDirection: Sony has not publicly committed to AV1.
- BoosteroidNow: H.264 primary, HEVC opt-inDirection: Slowest mover. AV1 not on the public roadmap.
- Amazon LunaNow: H.264 / HEVC depending on deviceDirection: AV1 expected via Twitch/AWS shared infrastructure in 2027.
What this means for you
On a fast, stable connection: the codec barely matters. You will not perceive the 5 ms latency difference, and you don't care about the bitrate.
On a marginal or capped connection: AV1 is a real win. If your data is capped at, say, 100 GB/month, AV1 effectively doubles how many hours you can cloud-game before hitting the ceiling.
For competitive shooters: prefer H.264 or HEVC. The ~3 ms saved at the encoder pipeline matters more than the bitrate when ranked points are on the line.
If you're shopping for a new device: ensure it has hardware AV1 decode. iPhone 15+ does. Most 2023+ Android flagships do. Apple Silicon Macs (M3 and later) do. Anything older falls back to software decode which adds 3–8 ms on top of the numbers above and burns battery.
Caveats and what we didn't measure
- VMAF is a perceptual metric tuned for video, not games. UI elements, sub-pixel font rendering, and rapidly moving HUDs can score better or worse than they look. We supplemented VMAF with blind A/B viewing tests on a 27" 4K display; results aligned.
- We tested at 1080p60. AV1's advantage grows at 4K and at 120 Hz — we'll publish a 4K-specific Labs piece later in 2026.
- The "end-to-end" column adds a constant 22 ms for server render + network transit. Your actual number depends on your distance to the data centre — see our region latency lab piece for that.