Video Codec ASIC Reference
01 Video Codec Standards the codec decides the file · the ASIC decides the speed & quality
| Codec | Year | Typical use | Compression vs H.264 | Status |
|---|---|---|---|---|
| H.264 / AVC MPEG-4 Part 10 | 2003 | Universal baseline · Blu-ray, Twitch ingest, low-latency live, every device made since 2008 decodes it | 1.0× (baseline) | Current the universal fallback |
| H.265 / HEVC | 2013 | 4K Blu-ray, Apple devices, 4K HDR streaming, OTT, security cameras · licence-heavy | ~50% smaller | Current mature · royalty cost real |
| VP9 | 2013 | YouTube primary codec · open-source · Google ecosystem · WebM container | ~45% smaller | Current YouTube backbone |
| AV1 AOMedia | 2018 | Royalty-free open standard · YouTube, Netflix, Twitch (rolling out), Discord streaming · the future | ~50% smaller | Current hardware encode mature since 2022 |
| VVC / H.266 | 2020 | Broadcast 8K, future OTT · software encoders only mostly · hardware decode starting in 2026 silicon | ~50% smaller than HEVC | Emerging 2026+ adoption |
| ProRes Apple | 2007 | Mezzanine / edit codec · large files, fast encode, near-lossless · cinema & post-production | 3–5× larger than H.264 | Current pro workflow standard |
| JPEG-XS | 2019 | Mezzanine / broadcast-IP transport · sub-frame latency · uncompressed-feel over SMPTE 2110 | ~5–10:1 visually lossless | Current broadcast-only niche |
| H.262 / MPEG-2 | 1995 | DVD, broadcast SD/HD, legacy capture | 3× larger | Legacy decode-only on modern hardware |
| VP8 | 2008 | WebRTC default video codec · obsolete elsewhere | ~same as H.264 | Legacy WebRTC fallback only |
02 NVIDIA NVENC / NVDEC on-die ASIC in every modern GeForce / Quadro / Tesla · separate from CUDA
| Architecture | NVENC gen | Year | Cards | Encoder capability |
|---|---|---|---|---|
| Pascal | 4th | 2016 | GTX 10-series, Tesla P4 | H.264 + HEVC 8/10-bit · no HEVC B-frames · 4K @ 60 |
| Turing | 6th | 2018 | RTX 20-series, GTX 16-series, T4 | HEVC B-frames · 8K @ 30 encode · ~15% better quality at same bitrate vs Pascal |
| Ampere | 7th | 2020 | RTX 30-series, A40, A10 | Same encoder as Turing · NVDEC adds AV1 decode |
| Ada Lovelace | 8th | 2022 | RTX 40-series, L40 / L40S | AV1 encode 8-bit · dual NVENC on 4080/4090 (40% faster) · L40S = 3× NVENC + 3× NVDEC |
| Blackwell | 9th | 2025 | RTX 50-series, B100, B200 | 4:2:2 chroma 10-bit H.264 + HEVC · MV-HEVC for stereoscopic / Vision Pro · improved AV1 UHQ mode · triple NVENC on 5090 |
NVDEC generations track behind. Key milestones: HEVC 10-bit on Pascal (3rd gen NVDEC), VP9 10-bit on Turing (4th), AV1 on Ampere (5th), 4:2:2 + improved AV1 on Blackwell (6th). All current cards decode every codec H.264 → AV1 at ≥ 4K @ 60.
03 Intel Quick Sync Video (QSV) on every Intel iGPU since 2011 · separate fixed-function block, free with the CPU
| Generation | Year | iGPU / dGPU | Encoder capability |
|---|---|---|---|
| Skylake → Coffee Lake Gen 9 / 9.5 | 2015–18 | HD 520 → UHD 630 | H.264 + HEVC 8-bit · HEVC 10-bit on Kaby Lake+ |
| Ice Lake / Tiger Lake Gen 11 / Xe-LP | 2019–20 | Iris Plus / Xe | VP9 encode · AV1 decode on Tiger Lake |
| Arc Alchemist Xe-HPG · DG2 | 2022 | Arc A380 → A770 | First AV1 encode on consumer hardware · 8-bit 4:2:0 · class-leading quality at the time |
| Meteor / Arrow Lake Xe-LPG / Xe2 | 2024–25 | Core Ultra 100 / 200 series iGPU | AV1 encode 8/10-bit · HEVC 4:2:2 10-bit |
| Arc Battlemage Xe2-HPG · BMG | 2024 | Arc B570 / B580 | AV1 + HEVC + H.264 enc/dec · improved low-latency for streaming · 8K @ 60 decode |
| Panther Lake Xe3 · Jan 2026 | 2026 | Core Ultra 300 iGPU | Full AV1 enc/dec 10-bit · HEVC 4:2:2 · early VVC/H.266 partial decode acceleration |
04 AMD VCN Video Core Next · replaced UVD/VCE on Vega/Raven Ridge
| VCN | Hardware | Notes |
|---|---|---|
| VCN 1.0 | Vega, Raven Ridge (2017) | H.264 + HEVC + VP9 decode |
| VCN 2.x | RX 5000 (RDNA 1) | HEVC encode improvements |
| VCN 3.0 | RX 6000 (RDNA 2), Ryzen 6000 APU | AV1 decode |
| VCN 4.0 | RX 7000 (RDNA 3), Phoenix | AV1 encode finally arrives · 8K @ 60 decode |
| VCN 5.0 | RX 9000 (RDNA 4, 2025) | Improved AV1 quality presets · matches NVENC Ada-class |
| VCN 5.0 (Strix Halo) | Ryzen AI Max+ APU | Full enc/dec all major codecs in a thin laptop |
05 Apple · Mobile · Datacenter specialist silicon outside the big three
| Engine | Codecs |
|---|---|
| Apple M1/M2/M3/M4/M5 AVE | H.264 + HEVC enc/dec · AV1 decode (M3+) · no AV1 encode even on M5 |
| Apple ProRes accelerator M-series Pro/Max | ProRes 422 / 4444 enc/dec hardware · M1 Max has 2× engines |
| Qualcomm Snapdragon X / X2 | H.264 + HEVC enc · AV1 decode · 8K @ 30 decode |
| NETINT Quadra T1U / T2A | Dedicated transcode ASIC · 32+ streams 1080p AV1/HEVC/H.264 per card · 7–15 W |
| AMD Alveo MA35D | Dual ASIC card · 32 channels AV1 transcode · for CDN / live streaming |
| Google Argos VCU | Internal YouTube transcoder · VP9 + H.264 + AV1 · not commercially available |
| NVIDIA L40S / L4 | Datacenter focus · L40S = 3× NVENC + 3× NVDEC · L4 = single-slot 72 W · most popular cloud encode card |
06 Codec × Hardware · Encode/Decode Support consumer / prosumer hardware as of mid-2026
| Codec / profile | NVENC Ada RTX 40 |
NVENC Blackwell RTX 50 |
QSV Battlemage Arc B580 |
QSV Xe3 Panther Lake |
VCN 5.0 RX 9000 |
Apple M5 AVE |
Snapdragon X2 Oryon Gen 3 |
|---|---|---|---|---|---|---|---|
| H.264 8-bit | |||||||
| HEVC 8-bit 4:2:0 | |||||||
| HEVC 10-bit 4:2:0 | |||||||
| HEVC 10-bit 4:2:2 | |||||||
| VP9 10-bit | |||||||
| AV1 8-bit | |||||||
| AV1 10-bit | |||||||
| VVC / H.266 | |||||||
| ProRes 422 | |||||||
| MV-HEVC (3D / Vision Pro) | |||||||
| JPEG-XS |
JPEG-XS is broadcast-only · all major consumer ASICs lack it · use NETINT, AMD Alveo, or Xilinx FPGA cards. ProRes encode is Apple-exclusive in hardware; others decode only via software/CPU.
07 Quick Picks match task → ASIC · current-gen consumer hardware
Game streaming · Twitch / YT
NVENC Ada / Blackwell
AV1 on Twitch ingest now live · best low-latency quality.
OBS recording
NVENC HEVC 10-bit
8 Mbps HEVC ≈ 14 Mbps H.264 · saves disk + GPU time.
Archival / Handbrake batch
Arc Battlemage AV1
Best quality/bitrate of any consumer encoder. Use SVT-AV1 CPU if patient.
Pro video editing
Mac Pro / Studio M-series Max
Dual ProRes engines · 4:2:2 hardware · Resolve / Premiere fly.
Multi-cam capture box
NETINT Quadra T1U / T2A
Dedicated transcode ASIC · ~10 W per 1080p AV1 stream.
Cloud transcode at scale
NVIDIA L4 / L40S
Most software (FFmpeg, Bitmovin) is best-tuned for NVENC.
CDN live AV1
AMD Alveo MA35D
32 channels AV1 per dual-slot card · < 1 W per channel.
Apple Vision Pro / spatial
M5 AVE + NVENC Blackwell
MV-HEVC encode required · only these two exist today.
Maximum quality, time-irrelevant
CPU SVT-AV1 / x265
Always beats ASIC at same bitrate. Just slow.
08 Critical Gotchas the small print that derails a pipeline
Hardware encode is fast, not optimalAt equal bitrate, NVENC / QSV / VCN output is ~10–20% larger than CPU x264/x265/SVT-AV1 at “veryslow” preset. ASIC is for live, real-time, batch throughput. CPU is for archive masters where quality-per-byte rules.
NVENC session limit was real, now liftedNVIDIA capped consumer GeForce cards at 3 simultaneous encode sessions via driver for years. The cap was removed in 2023 — current drivers permit unlimited sessions on RTX 30 / 40 / 50. Older articles still cite the old limit.
“AV1 support” is two different thingsDecode arrived on Ampere / Tiger Lake / RDNA 2 in 2020. Encode arrived only in 2022 (Ada, Arc Alchemist) and 2023 (RDNA 3). Apple still has no AV1 encode hardware in 2026. Verify which side of the codec you actually need.
HEVC royalties are a real billHEVC has three patent pools (MPEG-LA, HEVC Advance, Velos Media) plus unpooled patents. Commercial use fees can run into millions. AV1 is royalty-free by design — major reason for the data-centre shift.
4:2:0 vs 4:2:2 is a hard cliff for prosConsumer hardware historically only encodes 4:2:0 (consumer subsampling). Broadcast and pro acquisition use 4:2:2. NVENC Blackwell + Intel Arc Battlemage finally added 4:2:2 encode — game-changer for ENG / studio workflows.
B-frames not always supportedMaxwell HEVC, early VCN, and some mobile encoders skip HEVC / AV1 B-frames → lower compression efficiency (~10–15%). Check the encoder profile, not just “supports HEVC”.
Dual NVENC ≠ 2× throughput automaticallyAda RTX 4080/4090 and Blackwell 5090 have dual / triple encoders, but most apps (OBS, FFmpeg until recent builds) use only one. Need
-multi_encoder flag or NVENC split-frame mode. Same for L40S.Apple ProRes encode is Apple-exclusiveNo Windows / Linux machine encodes ProRes in hardware. Software encoding (FFmpeg) works but is slow and quality-locked. Pro shops doing ProRes deliverables run Mac Studio / Pro for that step alone.
# recommended
Login
0 Comments
Oldest