Tokens per watt, ranked

Speed leaderboards are everywhere. This one ranks qwen3-8b at Q4_K_M by how many tokens each machine produces per watt it draws. For a machine that lives on your desk, watts are the constraint: they set the heat, the heat sets the fan speed, the fan speed sets the noise, and the total decides whether your power supply has headroom. A card 20% faster that draws twice the power is not a better desk machine — it is a worse one that finishes sooner.

What this ranking rests on: 1 measured of 55 rows. 10 rows are modelled on an architecture whose efficiency constant was fitted from real runs; 44 rest on a constant nobody has calibrated. Every row below says which it is.

Hardware	Speed	Power	Tokens/watt	Evidence
1. Apple M5 Max	51 tok/s	35 W SoC package	1.50	ESTIMATED · uncalibrated
2. Apple M4 Max	46 tok/s	35 W SoC package	1.30	ESTIMATED · uncalibrated
3. Apple M1 Max	35 tok/s	30 W SoC package	1.20	ESTIMATED · uncalibrated
4. Apple M1 Ultra	64 tok/s	60 W SoC package	1.10	ESTIMATED · uncalibrated
5. Apple M2 Ultra	64 tok/s	60 W SoC package	1.10	ESTIMATED · uncalibrated
6. Apple M3 Ultra	65 tok/s	60 W SoC package	1.10	ESTIMATED · uncalibrated
7. Apple M3 Max	35 tok/s	35 W SoC package	1.00	ESTIMATED · uncalibrated
8. Apple M5 Pro	28 tok/s	30 W SoC package	0.93	ESTIMATED · uncalibrated
9. Apple M2 Max	35 tok/s	40 W SoC package	0.88	ESTIMATED · uncalibrated
10. Apple M4 Pro	25 tok/s	30 W SoC package	0.83	ESTIMATED · uncalibrated
11. Apple M5	14 tok/s	23 W SoC package	0.63	ESTIMATED · uncalibrated
12. Apple M1 Pro	19 tok/s	30 W SoC package	0.62	ESTIMATED · uncalibrated
13. Apple M2 Pro	19 tok/s	30 W SoC package	0.62	ESTIMATED · uncalibrated
14. Apple M4	11 tok/s	22 W SoC package	0.52	ESTIMATED · uncalibrated
15. NVIDIA H100 80GB	181 tok/s	350 W board	0.52	ESTIMATED · uncalibrated
16. NVIDIA A100 80GB	146 tok/s	300 W board	0.49	ESTIMATED · provisional fit (n=1)
17. Apple M2	10 tok/s	20 W SoC package	0.48	ESTIMATED · uncalibrated
18. Apple M3 Pro	14 tok/s	30 W SoC package	0.48	ESTIMATED · uncalibrated
19. Apple M3	10 tok/s	22 W SoC package	0.44	ESTIMATED · uncalibrated
20. NVIDIA GeForce RTX 5060	59 tok/s	145 W board	0.41	ESTIMATED · uncalibrated

On the SoC marker: unified-memory systems (Apple Silicon, DGX Spark, Strix Halo) report whole-package power — CPU, GPU and memory controller on one die — where a discrete card reports board power for the card alone. Neither figure is the wall draw of the complete machine. They are not the same measurement, so read the ranking as a real efficiency difference that is partly a difference in what is being counted.

How this is calculated. tokens_per_watt = decode_tok_s / watts. Watts resolve in one order and one only: a measured draw with its method where a run reports one, otherwise the hardware's spec power — board power for a discrete card, SoC package power for a unified-memory system — otherwise nothing at all. We never estimate a power figure for hardware that lacks one.

Also ranked: cost per 1M generated tokens, performance vs cost vs power vs VRAM, or the full filterable database.