Status: ReleasedData ownership: Institute-ownedLicence: CC BY 4.0
Silicon q4 throughput curves (tok/s vs config)
Apple-Silicon q4 decode throughput across configurations, behind the single-stream and campaign reports.
Purpose
Record decode throughput so runtime-planning claims can be checked against real numbers.
Source
Measured on the Apple-Silicon q4 path across the campaign and the single-encoder result.
Composition
Downloadable CSV — throughput (tok/s) across batch sizes B=1,2,4,8 for two configs (as-found and a setup-cache port), with vs-serialized ratios.
Known biases
- Warmup-inclusive and steady-state numbers differ materially (~11 vs ~24 tok/s).
Limitations
- Single-run point measurements with no reported variance.
- Hardware/toolchain versions not fully recorded.
Access conditions
The measured throughput table is available for download below, under CC-BY-4.0.
| config | batch_size | tok_per_s | vs_serialized | mode |
|---|---|---|---|---|
| as-found | 1 | 20.4 | 0.80 | batching |
| as-found | 2 | 28.0 | 0.80 | batching |
| as-found | 4 | 29.0 | 0.80 | batching |
| as-found | 8 | 33.0 | 0.80 | batching |
| setup-cache-port | 1 | 28.0 | 0.94 | batching |
| setup-cache-port | 2 | 32.3 | 0.94 | batching |
| setup-cache-port | 4 | 35.0 | 0.94 | batching |
| setup-cache-port | 8 | 35.6 | 0.94 | batching |