Mobile Video Recording: H.265, AV1 and ProRes Codec Analysis

The codec you choose now matters more than the sensor. Modern mobile SoCs can capture 8K30, but the entropy coding, motion estimation, and thermal budget behind H.265, AV1 and ProRes decide whether that footage is usable, shareable, or editable. Here is how they actually work on the silicon.

For three generations, H.264 was the default. It was simple, universally decodable, and thermally forgiving. That era is over. With 200MP sensors, 10-bit HDR, and platform mandates for efficient streaming, the encoder block has become a primary differentiator in flagship smartphone component sourcing decisions. Apple, Qualcomm and MediaTek now ship dedicated hardware for all three codecs — but they optimize for very different outcomes.

1. The Mobile Encoding Pipeline: Shared Constraints

All three codecs sit at the end of the same imaging chain. Light hits the sensor, the Image Signal Processor (ISP) performs demosaicing, noise reduction and tone mapping, and a 10-bit 4:2:0 frame is handed to the video encoder subsystem. That subsystem is where physics bites.

Real-time encode at 4K60 requires processing ~498 million pixels per second. At 8K30, it's ~995 Mpix/s. The encoder must perform motion estimation, transform, quantization and entropy coding within a ~16ms frame budget while staying under 2.5–3.2W for the entire media subsystem. Exceed that and the SoC throttles, dropping frames or forcing a codec fallback.

This is why on-device AI acceleration and encoder scheduling share the same NPU and memory bandwidth pool. Aggressive temporal noise reduction improves compression but also burns power before the encoder even starts.

2. H.265 / HEVC: The Efficient Workhorse

H.265 (High Efficiency Video Coding) remains the most balanced mobile codec. It improves on H.264 by ~50% bitrate savings using larger Coding Tree Units (64×64), 35 intra-prediction modes, and Advanced Motion Vector Prediction (AMVP).

How it works on mobile

Modern implementations (e.g., Qualcomm Spectra ISP + Adreno Video Encode) use fixed-function HEVC Main10 hardware. Key blocks:

• CTU partitioning: Recursively splits 64×64 blocks to 8×8 for detail preservation. Hardware does this in a single pass with lookahead of 2–3 frames.
• SAO & DEBLOCK filtering: Sample Adaptive Offset reduces banding in 10-bit HDR — critical for LOG and Dolby Vision profiles.
• WPP & Tiles: Wavefront Parallel Processing lets the encoder use 4–8 cores without breaking entropy sync, keeping latency to 25–35ms.

Typical rates: 4K30 at 60–80 Mbps (Main10, QP 22–26) is visually transparent for most smartphone optics. 4K60 pushes to 90–120 Mbps. 8K30 sits at ~160–200 Mbps.

H.265's strength is ecosystem and hardware maturity. Decoding is guaranteed on every device shipped since 2017, and hardware decode power is ~180mW for 4K60. For platform security, its bitstream parsers are also the most hardened — a factor noted in mobile kernel exploit mitigation reviews where malformed HEVC headers were a historic attack surface.

3. AV1: Compression King With a Power Cost

AV1, developed by the Alliance for Open Media, offers the best compression of the three — 20–30% over H.265 and up to 50% over H.264 in VMAF-matched tests (Netflix, Meta 2024 datasets). It achieves this with tools that are computationally brutal:

• 128×128 superblocks with 10-way partitioning
• 56 directional intra prediction modes plus palette and intraBC
• Warped motion and overlapped block motion compensation
• CDEF + Loop Restoration filters (replacing SAO)

On mobile, software AV1 encode is impractical (4–6x realtime on CPU). The shift happened with dedicated AV1 encoders in Dimensity 9300 and Snapdragon 8 Gen 3. These implement a constrained profile: 4K/8K, 10-bit, but with search range limited to ±64 pixels and only 2 reference frames to stay within thermal limits.

Result: AV1 at 4K30 needs only 40–55 Mbps for the same VMAF as H.265 at 70 Mbps. For streaming, that's decisive — YouTube and Netflix already prefer AV1 and will re-transcode uploads. But encode latency is higher (45–70ms) and file seeking is slower due to larger GOPs (up to 120 frames vs 30–60 for H.265 on mobile).

There is also a sensor angle. AV1's film grain synthesis preserves texture that H.265 smooths away, which benefits skin tones captured by modern health-integrated sensors that share the ISP pipeline. For archival content, AV1 is the longevity bet — royalty-free and increasingly hardware-decoded (all 2024+ flagships decode AV1 at ~210mW for 4K60).

4. ProRes: Editing-First, Not Delivery

Apple's ProRes (422 HQ, 422, LT, Proxy) is philosophically opposite. It is an intra-frame only, DCT-based intermediate codec. No inter-frame prediction. Every frame is independently coded with adaptive quantization.

ProRes 422 HQ at 4K30 is ~707 Mbps (Apple spec: 730 Mbps for 4K30p, 10-bit). ProRes 422 is ~500 Mbps. Compare that to 70 Mbps for H.265 — it's 7–10x larger. ProRes RAW is even higher, but is not generated on iPhone without external recorder.

Why tolerate it? Generation loss. Re-encoding H.265 or AV1 five times in an edit degrades chroma and introduces drift. ProRes survives 10+ generations with <0.2dB PSNR loss. Its constant-bitrate, I-frame structure scrubs at full speed in Final Cut and DaVinci, and hardware decode is trivial (Apple A17 Pro / A18 Pro have ProRes accelerator). Log recording (Apple Log) is only available in ProRes or HEVC+Log, and colorists prioritize ProRes for 10-bit 4:2:2 fidelity.

Constraints: Storage and interface. One minute of 4K60 ProRes HQ is ~6.1GB. iPhone 15/16 Pro limits ProRes 4K60 to external USB-C SSD for this reason, with internal recording capped to 4K30. Thermal is also higher — ~2.9W sustained due to massive memory bandwidth (ProRes writes at ~88 MB/s continuous).

5. Head-to-Head: Specs and Real-World Numbers

*With dedicated AV1 fixed-function block. Software fallback is >4W and not sustainable.

6. Developer & Creator Specs

Query capabilities at runtime — do not assume AV1 encode by OS version.

Bitrate math for planning: File size (GB) = (Bitrate Mbps × Duration sec) / 8192. A 10-minute 4K30 interview at H.265 80 Mbps = 5.86 GB, AV1 50 Mbps = 3.66 GB, ProRes HQ 707 Mbps = 51.8 GB.

For HDR, all three support 10-bit, but only H.265 and AV1 carry HDR10+ or Dolby Vision dynamic metadata efficiently. ProRes stores HDR as PQ/HLG with LUTs, which inflates grade time but preserves highlight roll-off better.

7. Choosing in 2026

There is no single winner — there is a pipeline decision:

• Everyday 4K60 / social: H.265 Main10 with HDR. Best compatibility, lowest thermals, instant sharing. Keep GOP at 1.5 sec for scrubbing.
• Maximum quality for upload: AV1 if your SoC has HW encode and your platform supports AV1 ingest (YouTube, Meta, Vimeo). Expect 25% smaller uploads at same VMAF, but monitor thermals — use a case off and avoid direct sun.
• Cinematic log & post: ProRes 422 (not HQ) on iPhone 15 Pro+ to external SSD when you will grade. Finish in H.265 or AV1 for delivery; never deliver ProRes to consumers.

Expect convergence. AV1 hardware encode power will fall below H.265 by 2027 as 3nm encoder tiles improve, and H.266 (VVC) mobile encoders are already sampling but not yet productized due to patent complexity. Until then, treat H.265 as default, AV1 as quality-efficient, and ProRes as an editing format that happens to be captured on a phone.

Next reads: Compare SoC scheduling trade-offs in our on-device LLM analysis and thermal impact on sustained recording in our BOM cost study. For security reviewers, see our HEVC bitstream hardening notes.