211 ns on the Raspberry Pi 5

We measured the hardware-vector-to-eBPF-entry latency on a Pi 5 at 211 ns average, 351 ns worst case, across 100 samples. The number is fixed: jitter sits at single-digit nanoseconds. Linux on the same board, measured with cyclictest, comes in at ~2 µs average and 7 µs worst.

Why we believe these numbers:

• CPU frequency scaling is disabled; the A76 is pinned at 2.4 GHz.
• The capture uses the Pi Debug Probe over UART0 at 115 200 baud — entirely out-of-band from the kernel under test.
• We’re measuring the path the GIC takes from a peripheral interrupt to the BPF program entry, with timestamps grabbed inside the architectural vector before any unrelated work runs.
• The kernel image itself is 10 MB, total footprint ~22 MB, so there is very little to interleave with the interrupt path in the first place.

The reason this is interesting is not the number alone — RTOSes have hit single-digit microseconds for years. It is that the path ends in eBPF, which we previously needed Linux for, and which we previously had to pay millisecond-class jitter for.

Full methodology and tables on the benchmarks page.