Whisper.cpp
Benchmark
The following comparison uses the latest official matrix of lzc_thor from 2026-03-21. The main table retains only models that have comparable results on all three platforms: Orin / Thor T4000 / Thor T5000.
| Model | Scene | Orin | Thor T4000 | Thor T5000 | Notes |
|---|---|---|---|---|---|
base.en | jfk, concurrency=16 | 3.39 req/s / P95 4.55 s / 56.82 W | 7.39 req/s / P95 2.10 s / 11.65 W | 9.44 req/s / P95 1.64 s / 17.68 W | All three devices can stably complete stress testing and are suitable as baseline for lightweight English language models. |
small.en | jfk, concurrency=16 | 2.46 req/s / P95 6.28 s / 26.88 W | 4.07 req/s / P95 3.80 s / 14.25 W | 5.36 req/s / P95 2.89 s / 21.76 W | More balanced between speed and recognition quality compared to base.en |
medium.en | jfk, concurrency=16 | 1.34 req/s / P95 11.52 s / 35.50 W | 1.77 req/s / P95 8.71 s / 16.85 W | 2.53 req/s / P95 6.11 s / 26.94 W | Advantages of the medium-sized model are more evident on Thor T5000 |
More
| Model | Scene | Thor T4000 | Thor T5000 | Notes |
|---|---|---|---|---|
large-v3-turbo | jfk, concurrency=16 | 1.59 req/s / P95 9.70 s / 20.93 W | 2.52 req/s / P95 6.13 s / 36.15 W | - 2026-03-21 The Orin results in the latest matrix are still incomplete and have not fully been recorded. Retaining only the Thor-to-Thor comparison for now. |
Notes
- Data source:
/home/catdog/lzc_thor/docs/runs/2026-03-21T08-45-00Z - The main table uses the official stress matrix results and does not include interrupted or unsynchronized cases.
- The power consumption column records the maximum power during testing, not the average power.