Opera:flags Experiments May 2026
Under the Hood: An Experimental Analysis of Feature Flags and Performance Tuning via opera:flags
Parallel downloading is highly effective on high-bandwidth, low-latency networks. However, on congested or high-latency links (simulated 100ms RTT), parallel downloads actually reduced throughput by 22% due to TCP head-of-line blocking. 4.3 Zero-Copy Rasterizer ( #enable-zero-copy-rasterizer ) | Metric | Default (copy) | Zero-copy ON | Δ | | --- | --- | --- | --- | | Scrolling smoothness (MotionMark) | 812.5 | 798.2 | -1.8% (regression) | | First paint (ms) | 214 ms | 221 ms | +3.3% slower | | GPU Memory (MB) | 342 MB | 358 MB | +4.7% | | Stability | Stable | Crashes on YouTube (Hardware-accelerated video) | Fatal | opera:flags experiments
GPU rasterization significantly improves graphics-heavy benchmarks but nearly doubles GPU memory. No JavaScript performance regression. 4.2 Parallel Downloading ( #enable-parallel-downloading ) | Metric | Enabled (default) | Disabled | Δ | | --- | --- | --- | --- | | 100MB download time (sec) | 1.82 (±0.09) | 2.97 (±0.12) | +63% slower when disabled | | CPU during download (%) | 12% | 4% | -66% CPU when disabled | | Max TCP connections | 4 | 1 | N/A | Under the Hood: An Experimental Analysis of Feature
| Page tested | Crashes (out of 10 loads) | | --- | --- | | YouTube (any video) | 10 | | Vimeo | 8 | | Local MP4 | 9 | | Google Search | 0 | No JavaScript performance regression