Benchmarks
What one server carries, measured. Each table says how it was measured, and what was not.
The test server
Measured 2026-10-13 on creekX 2026.10.13, installed with the standard one-line installer, on a cloud server with 4 vCPU, 8 GB memory and 100 GB NVMe, Ubuntu 24.04, about 890 Mbit/s upload. Source: a real 1080p25 H.264 clip at 5 Mbit/s, looped. Each step ran 60 seconds. CPU figures are creekX's own; the test's senders and viewers are subtracted.
Incoming lives and cameras (1080p, each recorded)
| Lives at once | creekX CPU (of 4 cores) | Memory | Disk writes |
|---|---|---|---|
| 10 | 5% | 0.8 GB | 3.7 MB/s |
| 25 | 7% | 1.1 GB | 9.3 MB/s |
| 40 | 10% | 1.4 GB | 17.6 MB/s |
| 49 | 13% | 1.5 GB | 30.6 MB/s |
All stayed live. creekX records without re-encoding, so the limit is incoming bandwidth and disk, not the processor: a 1080p stream at 5 Mbit/s writes about 2.3 GB an hour. The test stopped at 49 because a workspace holds 50 streams, not because the server was full.
Real-time (WebRTC) viewers of one live
| Picture | Bitrate | Viewers, all receiving | creekX CPU |
|---|---|---|---|
| 1080p | 2,500 kbit/s | 500 | 37% |
| 720p | 2,000 kbit/s | 400 | 39% |
| 480p | 1,000 kbit/s | 800 | 47% |
| 360p | 800 kbit/s | 800 | 47% |
The viewers ran on the same server, so the test itself ran out first. Off the server the network port is the limit: on about 890 Mbit/s that is roughly 340 viewers at 1080p, 430 at 720p and 850 at 480p (calculated, not measured).
Low-latency HLS viewers of one 1080p live
| Viewers | creekX CPU | Delivered | Stalled |
|---|---|---|---|
| 100 | 0.29 cores | 489 Mbit/s | 0 |
| 200 | 0.44 cores | 1,013 Mbit/s | 0 |
| 400 | 0.80 cores | 2,258 Mbit/s | 0 |
Several qualities (720p, 480p, 360p from 1080p)
About 0.6 of a core per live. The 4-core server made qualities for 4 lives comfortably (64% CPU); 6 was too many.
Delay behind live
| Way of watching | Median | Worst |
|---|---|---|
| WebRTC | 0.49 s | 0.68 s |
| Low-latency HLS | 1.23–1.28 s | 1.40 s |
| HLS (several qualities) | 3.4–3.8 s | 4.4 s |
Measured 2026-10-08 on a laptop over the local network, OBS sending, tuned for real time.
AV1 copies of library videos
At the same picture quality (VMAF), AV1 copies were 25–30% smaller than H.264; one or two VMAF points lower, 40–65% smaller. Measured 2026-10-12 on an 8-core machine, FFmpeg 9.0.2, on 720p and 480p test films.
Not measured yet
- 24/7 channels per server (each channel used 4–6% of a core in development).
- Viewers on many real networks, and servers larger than 4 cores.
- Real IP cameras over days of recording.
Size your own deployment with the pricing calculator, or try it free on your hardware.