Our Measurement Methodology

Internet speed is not a single static number but a distribution of measurements that depends on time and conditions. Beyond bandwidth, NQF presents responsiveness under load (RPM, bufferbloat) and which digital experiences you can enjoy smoothly with your connection as a first-class output.

Download is done over an HTTP/2 stream and upload primarily with WebSocket binary frames (with an HTTP/2 fallback if WebSocket cannot be established). Latency is measured over a separate connection from the download/upload traffic, so latency under load is not distorted. The server is chosen by the lowest median RTT and test traffic is pinned to a single host.

The reported speed is the P30–P90 trimmed mean obtained by first reducing the samples collected over the 15-second test into 20 buckets and then discarding the bottom 30% and top 10%. This eliminates burst spikes and gives the sustainable real speed.

Test duration

15 s fixed (download + upload)

Download transport

HTTP/2 streaming GET

Upload transport

WebSocket (primary) · HTTP/2 (fallback)

Download concurrency

clamp(round(Mbps/6), 4, 6) streams

Upload concurrency

4 WS sockets · 64 KB chunk · 250 ms bucket

Reported speed

bucketize(20) → P30–P90 trimmed mean

Server selection

Lowest median RTT · single-host pinning

Ping

5 samples/node · 50 ms · minimum RTT

RPM

min(round(60000 / loadedLatency), 2000) · RFC 9097

Bufferbloat

max(0, latency under load − idle ping)

4K Video Streaming

Download > 25 Mbps

Cloud Gaming

RPM > 800 · Download > 30 Mbps · ping < 40 ms

Video Call

RPM > 200 · ping < 150 ms

Our engine is calibrated to share the same core statistical reporting (15 s test, P30–P90 trimmed mean, clamp(round(Mbps/6),4,6) concurrency) as the industry's global reference, Speedtest.net v2.11. The points that differ by design — WebSocket-first upload and single-host topology — are explicitly listed in the whitepaper. In an independent three-run comparison, the download deviation was measured at the ±0.2% level.

The following methods are design goals; they are not part of the current engine: distribution modeling (KDE) and bimodality detection, spectral analysis with the Welch periodogram, path decomposition, HTTP/3 + QUIC transport, Z-score outlier handling and native (iOS/Android) clients.