TECHNICAL
Resolution
VMAF (Standard)
VMAF (w/Optimizer)
Improvement
1080p
94.08 78.61 62.47 51.25 45.76
94.62 92.20 85.35 66.59 58.25
+0.54
720p 540p 432p 360p
+13.59 +22.88 +15.34 +12.49
CONCLUSION Key Findings
n NVIDIA’s NVENC Excellence: The NVENC pipeline sees universal improvement, with VMAF gains of 6-15 points, pushing scores into the 90+ range. n Rescuing Low Bitrates: For NETINT and Intel hardware, the Optimizer acts as a safety net for lower-mid resolutions, often increasing VMAF scores by 15 to 29 points. This is crucial for maintaining viewer retention on mobile networks. n Software Efficiency: VisualOn brings software encoding (AVC/HEVC) closer to the theoretical maximum quality for 1080p/720p streams. For Cires21 customers, this integration translates directly to reduced churn, lower bandwidth requirements for equivalent quality, and a consistent viewing experience regardless of the end-user’s device or network conditions. Test Platforms n For NVIDIA’s NVENC: NVIDIA RTX 4000 Ada Generation. n F or NETINT’s VPU: NETINT Ouadra’s T2A n For all CPU-based tests: AMD EPYC 74F3 24-Core Processor For Intel’s OneVPL/OuickSync tests: The testbeds used have Intel Corporation Raptor Lake-P [ Iris Xe Graphics] integrated GPU and Intel’s Arc A77016GB.
ABR ladders at the boundaries of Pareto efficiency per encoder implementation, with and without the VisualOn Optimizer enabled.
Processing performance assesment NVIDIA’s NVENC: For our l:N ladder, we see an SM load increase of -10% as monitored by ‘nvidia-smi’ at full load. Throughput impact on NVENC is negligible. Intel’s OneVPL/0uickSync: For our l:N ladder, we see an VPP (Video Post Processing) engine load increase of -22% as monitored by ‘intel-gpu-top’ at full load. Throughput impact on OneVPL/OuickSync is only noticeable on more complex 1:N ladders, well beyond the scope of this paper, affecting the maximum simultaneous encode sessions per GPU. NETINT 0uadra’s VPUs: For our l:N ladder, we see an Al engine load increase of -80% as monitored by NETINT’s ‘ni_rsrc_mon’ tooling at full load. Each of the VPUs on the Ouadra T2A showed identical Al engine load, proportional to the encoding load in use, regardless of the best model load set at runtime. Throughput impact on NETINT is noticeable on all l:N ladders, especially as the Al engine load increases with the model(s) being loaded into the VPU.
And all tests were conducted on Cires2l’s Live Transcoder with VisualOn’s Optimizer integration.
Test content
For these tests, we used Netflix’s Meridian test footage - an ideal benchmark for evaluating encoder quality retention. Meridian is an open-source, mathematically hostile sequence engineered to stress-test video compression algorithms and perceptual quality metrics like VMAF. It aggregates severe spatial and temporal complexities - high-frequency film grain, volumetric fog, unpredictable motion vectors, and 4,000-nit HOR contrast extremes - making its 4K 60fps mezzanine files a pristine ground- truth reference for full-reference metric comparisons. Access to the uncompressed masters allowed our R&D team to rigorously analyse rate control behaviour, expose macroblocking or chroma banding, and accurately map the convex hull for
www.cires21.com | www.visualon.com
Volume 48 No.23 SEPTEMBER 2026
107
Made with FlippingBook - Online magazine maker