1. Pin the workload
Use the same composition, frame range, resolution, frame rate, codec intent, and input bytes. Name the suite so a later run can be compared.
Benchmark methodology · updated 2026-07-23
How we compare Mac rendering with cloud runs.
A benchmark is useful only when the workload, host, measurement, and claim boundary are visible. RenderMac publishes the same composition or source through each available path, reports wall-clock first, and separates measured results from projections, rate-card estimates, and package-power samples.
Current source set: the 2026-07-22 Remotion fleet, long-kinetic cloud comparison, VideoToolbox encode, and M2 Max
powermetrics runs in research/apple-silicon-idle-render-network/benchmarks/results/. The benchmark pages are evidence summaries, not invoices or a promise that every cloud shape behaves the same way.Five rules for a fair render benchmark
2. Record the whole wall-clock
Include the render and encode path that a buyer waits for. Call out cold-start, scaffolding, or transfer steps instead of hiding them.
3. Identify the machine
Publish chip family, memory, OS/runtime, runner version, container CPU/memory shape, and whether the path was local, live cloud, or a projection.
4. Measure power honestly
Label package watts or watt-hours as on-device SoC estimates. Do not present them as wall-outlet measurements or idle-subtracted energy when the idle sample is missing.
5. Keep unlike scaling separate
A single Mac and a massively sharded Lambda fleet answer different questions. We compare the tested shape and explicitly say when a fan-out path was not run.
How to read our tables
- Measured: the job completed on the named host and the elapsed time is recorded.
- Projected: a transparent scale from a measured row, such as 20 minutes from a 3-minute cloud run; it is not a completed job.
- Rate-card estimate: published compute rates multiplied by measured wall-clock, not a RenderMac invoice or negotiated cloud bill.
- ABS package energy: absolute SoC energy that includes baseline draw when an idle-subtracted sample is unavailable; use it as an upper-bound context signal.
- Not comparable: a missing binary, permission boundary, or different execution model. We leave the cell blank rather than inventing a zero.
What the current tests do - and do not - show: M2 Max and M1 Airs were materially faster than the tested small Cloudflare Containers shapes for Chromium Remotion and the measured software encode. That does not prove a single Mac beats every cloud configuration, Remotion Lambda fan-out, or a larger dedicated instance. It shows where Apple Silicon is worth testing for a given workload.
Browse the measured runs
Remotion vs cloud
Same compositions on M2 Max, M1 Air, live Containers, and a Docker-matched shape.
M1 Air vs M2 Max
Fleet wall-clock comparison across LightMotion, MediumMotion, HeavyMotion, and a long kinetic check.
VideoToolbox vs x264
A 20-minute 1080p30 encode showing Mac-native VideoToolbox and software x264 beside the cloud-shaped run.
Power draw
M2 Max package watts, watt-hours, and the practical thermal implications for hosts.