Data
Every measurement, in one table.
Every energy figure behind our research. Measured, never fitted, each row traced to its note.
Download the CSV
Compare
Pick two, see the gap.
ratio-
tokens per kWh, A-
tokens per kWh, B-
from one phone battery (15 Wh), A / B-
Per correct answer
Energy per right answer.
| machine | model | format | calibration | at once | correct | J per correct | tok/s | note |
|---|---|---|---|---|---|---|---|---|
| Apple M5 Pro | 1.5B | Q4_0 | none | 1 | 66.0% | 40.2 | 197 | 19 |
| Apple M5 Pro | 1.5B | Q4_0 | none | 8 | 65.6% | 13.7 | 866 | 19 |
| Apple M5 Pro | 7B | Q4_0 | none | 1 | 88.4% | 101.9 | 60 | 19 |
| Apple M5 Pro | 7B | Q4_0 | none | 8 | 88.0% | 35.1 | 279 | 19 |
| AMD Ryzen AI Max+ 395 | 1.5B | Q4_0 | none | 1 | 66.0% | 230.9 | 178 | 19 |
| AMD Ryzen AI Max+ 395 | 1.5B | Q4_0 | none | 8 | 66.6% | 65.2 | 648 | 19 |
| AMD Ryzen AI Max+ 395 | 7B | Q4_0 | none | 1 | 88.8% | 609.0 | 48 | 19 |
| AMD Ryzen AI Max+ 395 | 7B | Q4_0 | none | 8 | 88.4% | 156.6 | 210 | 19 |
| Apple M5 Pro | 1.5B | Q2_K | none | 8 | 21.0% | 39.2 | 773 | 20 |
| Apple M5 Pro | 1.5B | Q2_K | imatrix, general text | 8 | 37.2% | 28.4 | 836 | 20 |
| Apple M5 Pro | 1.5B | Q2_K | imatrix, GSM8K-train | 8 | 42.4% | 23.8 | 818 | 20 |
| Apple M5 Pro | 1.5B | Q4_0 | none | 8 | 65.8% | 14.8 | 851 | 20 |
GSM8K, first 250 test questions. Chip energy. spliced-quality.csv.
The table
mJ per generated token. Each row names its measurement point. Method in its note.
Raw files
Everything else.
Every pass of every experiment is in phase/results as JSON, including the ones the notes discard and why. The scripts that produced them are in phase/experiments.
The data is free to reuse with attribution. See CITATION.cff.