SD.CPP

Benchmark

The data is based on the official lzc_thor result from 2026-03-21T03-10-55Z's speed_matrix; only models with valid results from all three Orin / Thor T4000 / Thor T5000 instances are retained.The boundary detection, single-service/double-service comparison, and smoke test cases are not included in this table.

Model	Best Case	Orin	Thor T4000	Thor T5000
FLUX.1-dev	Based on the best case for each machine	0.033 img/s / 30.72 s / 44.75 W	0.047 img/s / 42.73 s / 20.59 W	0.068 img/s / 29.41 s / 38.41 W
SD2.1	Based on the best case for each machine	0.158 img/s / 28.58 s / 48.54 W	0.183 img/s / 46.49 s / 20.16 W	0.252 img/s / 9.94 s / 32.42 W
SD3 Medium	Based on the best case for each machine	0.072 img/s / 20.83 s / 49.16 W	0.148 img/s / 13.48 s / 20.55 W	0.237 img/s / 8.44 s / 38.28 W
SDXL	Based on the best case for each machine	0.030 img/s / 67.45 s / 48.28 W	0.040 img/s / 50.10 s / 20.93 W	0.048 img/s / 41.34 s / 34.31 W
SSD-1B	Based on the best case for each machine	0.217 img/s / 11.56 s / 47.82 W	0.243 img/s / 16.44 s / 20.93 W	0.357 img/s / 11.21 s / 34.69 W

Best Case

Model	Orin Best Case	Thor T4000 Best Case	Thor T5000 Best Case
FLUX.1-dev	p1	batch_2	batch_2
SD2.1	p8	p16	p4
SD3 Medium	p2	batch_2	batch_2
SDXL	batch_2	batch_2	batch_2
SSD-1B	p4	batch_4	batch_4

Notes

• Data source: /home/catdog/lzc_thor/docs/runs/2026-03-21T03-10-55Z
• The second number in the table represents the average latency of the best case.
• The power consumption column records the maximum power during the test.
• FLUX.2-dev and FLUX.2-klein did not produce stable and usable results in the latest test, and thus are not included in the official benchmark table.