NETINT VPU ECOSYSTEM · IBC 2026
AT A GLANCE
This article explains how MPEG-5 LCEVC improves video efficiency by combining a lightweight enhancement layer with existing codecs such as H.264, HEVC, AV1, and VVC. It reviews published subjective studies, addresses common technical objections, and examines how LCEVC reduces bitrate while maintaining visual quality.
The article also explores why pairing MPEG-5 LCEVC with NETINT VPUs creates a more efficient streaming workflow by reducing both encoding costs and CDN bandwidth, helping video platforms optimize compute resources and delivery efficiency at scale.
Every video platform pays for scale twice. The first bill is processing: transcoding servers, cloud instances, power, and cooling. The second bill is delivery: CDN egress, cache efficiency, time to first picture, rebuffering, and the quality viewers see on constrained networks. Processing efficiency and delivery efficiency are two levers on the same scaling problem, and most teams pull only one.
NETINT addresses the processing layer with video processing units (VPUs), ASIC-based encoders that replace racks of CPU and GPU transcoders with highly efficient purpose-built video accelerators. V-Nova addresses the delivery layer with MPEG-5 LCEVC, a standardized enhancement codec that cuts bitrate at equal subjective quality. This post explains how LCEVC works, reviews published subjective evidence, and shows why the two technologies compound when deployed together.
How LCEVC works
MPEG-5 LCEVC (ISO/IEC 23094-2) is a standardized enhancement layer that sits on top of any base codec: H.264, HEVC, VVC, or AV1. The base codec encodes at reduced resolution, typically half the width and height, which is one quarter of the output pixels. Two lightweight residual sub-layers then restore full-resolution detail. The first corrects the decoded base picture before upsampling. The second adds the fine detail that the downscale removed, at full output resolution.
The enhancement travels as SEI metadata inside the base stream or as a separate CMAF track. Either way, packaging, CDN mechanics, and DRM stay largely untouched. A device without LCEVC player support ignores the enhancement and renders the base layer at its native resolution. In the Brazilian broadcast tests described below, the enhancement layer accounted for 9% to 14% of total bitrate; the base layer carried the rest.
That split is the point. The base layer is the expensive part of an LCEVC stream to encode, and it’s what a VPU accelerates.
What the subjective tests show
The LCEVC record rests on formal subjective testing, not objective metrics alone. Four rounds matter most.
- MPEG verification tests (2021)
MPEG’s verification test used the BT.500 DSIS method with 50 screened viewers across two independent laboratories. Against full-resolution single-layer encodes, LCEVC saved 45.9% for AVC at UHD and 28.5% at HD, and 30.9% for HEVC at UHD and 24.1% at HD, measured as MOS BD-rate. A second anchor tested the obvious alternative: the same codec at half resolution, upscaled by the decoder. LCEVC beat that anchor by 21% to 38% depending on codec and resolution (Figure 1).
Figure 1. MPEG verification tests: subjective BD-rate against two anchors. Data: ISO/IEC JTC 1/SC 29/WG 04 N0076 (2021), Tables 13 and 14. Charts recreated.
2. Brazil SBTVD Phase 3 (2024)
Brazil’s TV 3.0 project ran real-time encoding tests at the University of Brasília under BT.500-15, with 30 viewers per test. The Forum’s method sets the output bitrate for each configuration at the worst-case clip: the highest bitrate any of the five test clips needed to match a 1080p VVC reference. On that basis, single-layer 4K VVC needed 12.54 Mbps. VVC plus LCEVC reached the same 4K resolution and quality target at 8.08 Mbps, a 36% reduction.
Per clip, the gap ranged from 14% (globo05, a football sequence) to 75% (philips01), as Figure 2 shows. The Forum also flagged a confound: the two arms used different encoders, a 120-frame GOP for single-layer VVC against a 360-frame GOP for the LCEVC base. That caveat raised the right question, and it prompted the next round.
Figure 2. The consortium round that raised the right question. Data: SBTVD Forum Phase 3, February 2024, Tables 22 and 36. Grade 0 equals the 1080p reference. The Forum sets each arm’s output bitrate at its worst-case clip: philips01 for single-layer VVC (12.54 Mbps), globo05 for VVC plus LCEVC (8.08 Mbps). Confidence intervals overlap at adjacent points.
3. Matched-encoder retest (2024)
V-Nova commissioned the University of Brasília to repeat the 4K comparison with the confound removed: the same MainConcept live encoder and a 120-frame GOP in both arms, under the same protocol with at least 30 viewers per test. Single-layer 4K VVC needed 16.58 Mbps to match the reference. VVC plus LCEVC needed 7.79 Mbps, 53% less at equal subjective quality. Per-content savings ran from 46% to 53.5% across the five sequences (Figure 3).
Figure 3. Matched encoders: equal 4K quality at less than half the bitrate. Data: UnB Additional Tests for TV 3.0, August 2024, Tables 10 and 21. Error bars are published 95% confidence intervals.
4. Newer and independent results
Results published since then point the same way. V-Nova’s peer-reviewed QoMEX 2026 study of LCEVC-enhanced VVenC measured 23% bitrate savings at equal subjective quality (BD-MOS) with a 2.6x faster encode. A University of the West of Scotland preprint (January 2026) found LCEVC-enhanced VVC and multilayer VVC delivered comparable subjective quality, with both ahead of base-layer upsampling at the stronger operating point. Kingston University and TU Berlin measured a 42% VMAF BD-rate saving versus x264 and 39% versus x265 for 1080p60 gaming content. The University of Surrey measured 18.7% and 15.8% savings over H.264 and HEVC at CRF 19, and found the gains held under packet loss.
The standard has also moved into production. Brazil’s TV 3.0, branded DTV+, was established by presidential decree published on August 27, 2025, with LCEVC enhancing VVC as the video system. V-Nova reports commercial 4K HDR service under 10 Mbps in Rio de Janeiro, São Paulo, and Brasília since June 2026. FFmpeg (7.1), GStreamer (1.26), and Shaka Player carry LCEVC upstream, and ATSC A/345 lists LCEVC-based scalability as an option for NextGen TV.
| Study | Year | Method | Headline result | Independence |
|---|---|---|---|---|
| MPEG verification tests (WG04 N0076) | 2021 | BT.500 DSIS, 50 viewers, 2 labs | AVC −45.9% UHD / −28.5% HD; HEVC −30.9% / −24.1% (MOS BD-rate) | MPEG-supervised, independent labs |
| SBTVD Forum Phase 3 (Brazil) | 2024 | BT.500-15, 30 viewers per test | 4K parity at 8.08 vs 12.54 Mbps (−36%); encoders differed between arms | Government-funded, run by UnB |
| UnB retest, matched encoders | 2024 | BT.500-15, 30+ viewers per test | 4K parity at 7.79 vs 16.58 Mbps (−53%); per-content −46% to −53.5% | V-Nova commissioned, UnB executed |
| Kingston Univ. + TU Berlin (ACM MHV) | 2022 | VMAF BD-rate + P.809 panel | −42.1% vs x264, −38.9% vs x265, 1080p60 gaming | Independent; SDK from V-Nova |
| University of Surrey (Springer MTA) | 2024 | VMAF, SSIM, PSNR; packet loss | −18.7% (AVC) / −15.8% (HEVC) at CRF 19; reverses at CRF 27/35 on AVC | Independent |
| Univ. of the West of Scotland | 2026 | DCR, 25 viewers, 15 UHD clips | LCEVC ≈ multilayer VVC; both beat base upsampling at the stronger operating point | Independent; bitstreams from V-Nova |
| V-Nova VVenC study (QoMEX) | 2026 | DCR at independent facility, 24 viewers + metrics | BD-MOS −23%; VMAF −30%; PSNR +77%; 2.6× faster encode | Vendor-authored, peer-reviewed venue |
| MediaTek (Picture Coding Symposium) | 2024 | PSNR under JVET CTC + expert viewing | PSNR losses vs VTM/HM references; no clear visual benefit seen at 4K | Independent, adversarial |
| Leibniz Univ. Hannover (ICASSP) | 2025 | PSNR-domain BD-rate sweep | Optimized LCEVC still trails full-res VTM by ≈15% (PSNR) | Independent, adversarial |
Table 1. The published record, including adversarial results. Sources are listed at the end of this article.
Engineers who have watched layered codecs fail before are right to be skeptical. Here are the five objections that come up most, with what the record says about each.
“Layered coding lost this argument once already. SVC and SHVC went nowhere.”
They did. SVC added overhead and complexity, and no ecosystem carried the cost. LCEVC inverts each failure mode. The enhancement is 9% to 14% of the stream, not a parallel prediction structure. Encoding gets cheaper, not more expensive: 2.6x faster for VVenC and roughly 40% fewer SVT-AV1 cycles at equal VMAF-NEG quality (SPIE, 2022). Legacy decoders play the base layer. SVC’s costs landed before its benefits. LCEVC runs the other direction.
“The gains are a metric artifact. Show me PSNR.”
Partly true. LCEVC optimizes perceptual reconstruction, and PSNR punishes it. The same VVenC encodes that saved 23% with human viewers and 30% on VMAF lost 77% on PSNR (Figure 4). VMAF correlated with those viewers at 0.89; PSNR correlated at 0.49. MediaTek (PCS 2024) found PSNR losses against reference VTM and HM encoders and no clear visual benefit at 4K. Leibniz University Hannover (ICASSP 2025) found that even optimized LCEVC trails full-resolution VTM by roughly 15% in PSNR-domain BD-rate.
Both adversarial studies use reference encoders, operate above delivery bitrates, and score the one metric LCEVC does not optimize. With human viewers at delivery bitrates, the LCEVC record is consistent and validated. Teams that gate codec decisions on PSNR should expect LCEVC to fail that gate, and should ask whether the gate measures what their viewers see.
Figure 4. Same encodes, five verdicts: the metric decides the argument. Data: Razaak, Tahir, and Cobianchi (V-Nova), QoMEX 2026, Tables II to IV.
“This is upsampling with extra steps.”
MPEG tested exactly this as its second anchor: the same codec at half resolution, upscaled. LCEVC beat it by 21% to 38% depending on codec and resolution (Figure 1, right panel), and the West of Scotland study repeats the finding for VVC. Ladders that drop resolution have the right instinct. LCEVC adds a standardized channel for the detail the downscale threw away.
“Nothing wins everywhere.”
Three boundaries are published. Operating point: Surrey’s AVC gains at CRF 19 reversed at CRF 27 and 35, and enhancement layers near 10% of total bitrate sometimes landed within the confidence interval of plain upsampling in the West of Scotland data. Content: per-clip savings in Brazil’s Phase 3 round ranged from 14% to 75%, and from 46% to 53.5% in the matched-encoder retest. Metric: PSNR-domain evaluations against reference encoders show LCEVC behind, as the previous section describes.
“How independent is this evidence?”
The QoMEX study is V-Nova-authored. The UnB retest was V-Nova-commissioned. The academic studies used V-Nova’s SDK, since no other production implementation exists. Against that: the verification tests ran under MPEG supervision in two independent labs, Brazil’s Phase 3 was government-funded, and the field publishes adversarial results. Formal subjective tests at delivery-relevant operating points show 23% to 53% savings over five years. It is right to demand public protocols. It is also a mistake to ignore replicated results across disparate studies.
Two layers, two proofs
The pairing with VPUs is arithmetic. The base layer is the expensive part of an LCEVC stream, and the VPU accelerates it. Silicon that encodes N 4K channels encodes roughly 4N 1080p base layers, because the base runs at one quarter of the output pixels.
Processing efficiency helps platforms create more streams with fewer resources. Compression efficiency helps them move those streams with fewer bits. Workflow efficiency decides whether either improvement can be used in production.
Encoding-efficient hardware that also delivers bandwidth savings belongs at the top of every video team’s technical agenda. Platforms that chase bitrate savings alone will struggle to produce efficient outputs economically. Platforms that chase compute savings alone give the gains back on the CDN invoice. The stronger path evaluates both layers, compute efficiency and bandwidth efficiency, and holds each to the same priority.
Summary
MPEG-5 LCEVC reduces bitrate at equal subjective quality by 23% to 53% in formal viewer tests, with the enhancement layer carrying 9% to 14% of the stream. The base layer, which carries the rest, runs at one quarter of the output resolution and is exactly the workload a VPU accelerates. The adversarial results are real, and they measure PSNR against reference encoders above delivery bitrates. For teams evaluating both technologies, the test that matters is a subjective one at your delivery operating points, on your content, with your encoder in both arms.
NETINT VPU ECOSYSTEM · IBC 2026
As video workloads continue to grow, adding capacity is no longer a single decision, but a set of practical paths depending on the real bottleneck. This series explores how to scale video processing efficiently using NETINT VPUs, whether through infrastructure partners, existing systems, or dedicated deployments designed around video from the start.
1. The New Video Capacity Problem
2. Three Ways to Add Video Capacity in a Constrained Market
3. VPU-Powered Capacity as a Service
4. Get More From the Servers You Already Own
5. Build Dedicated VPU-Based Video Infrastructure
6. VPU Capacity Without a Hardware Cycle
7. Scaling Performance-Critical Video Workloads
8. Beyond General-Purpose Compute
9. Reliable Live Video Meets Efficient Processing
10. Maximizing Encoding Efficiency in Live Streaming Workflows
11. Pairing VPUs with MPEG-5 LCEVC Delivers The Next Layer of Video Efficiency
12. From Concept to Production: A Field-Proven Methodology for Deploying NETINT Quadra VPUs
Subscribe to the NETINT blog
One email a week with our latest articles on VPUs, encoding efficiency, and video infrastructure.
