Edge AI measurements: Jetson AGX Orin 110 tok/s @ 37W, 3 tok/s per watt record
10-16 batch 108 edge / mobile device measurements: Jetson AGX Orin 64GB avg 110.96 tok/s (37.5W), Snapdragon X Elite 80.82 (22.5W), Snapdragon 7 Gen 4 78.59 (21.2W). **Per-watt tok/s: Jetson 2.96 — DB record**.
Source: araoai
Edge AI measurements: Jetson AGX Orin 110 tok/s @ 37W, 3 tok/s per watt record
10-16 batch has 108 edge / mobile device L2 records — Jetson + Snapdragon mobile chips first large-sample in DB.
Overall recap
| Hardware | VRAM | Records | Avg tok/s | Avg power | Per-watt tok/s |
|---|---|---|---|---|---|
| Jetson AGX Orin 64GB | 64GB | 45 | 110.96 | 37.5W | 2.96 |
| Qualcomm Snapdragon X Elite | 16GB | 36 | 80.82 | 22.5W | 3.59 |
| Qualcomm Snapdragon 7 Gen 4 | 8GB | 27 | 78.59 | 21.2W | 3.71 |
| Ref: RTX 5090 desktop | 32GB | 35 | 270.04 | 450W | 0.60 |
| Ref: Apple M5 Max | 128GB | 34 | 257.67 | 145W | 1.78 |
Key observations:
- Snapdragon 7 Gen 4 at 3.71 tok/s/W is DB highest — phone-class chip running LLM is finally practical
- Jetson AGX Orin 110.96 tok/s @ 37.5W = 2.96 tok/s/W — 4.9× higher than desktop RTX 5090 (0.60 tok/s/W)
- Jetson 64GB fits 27B IQ2_XS — edge device running big models is no longer "barely"
Jetson AGX Orin 5 most dramatic single records
| Model | Quant | tok/s | VRAM | power |
|---|---|---|---|---|
| Qwen3.5-0.8B | IQ1_S | 337.39 | 2.80 GB | 32W |
| Qwen3-1.7B | Q3_K_S | 215.20 | 4.10 GB | 35W |
| Qwen3.5-4B | Q3_K_M | 156.50 | 7.50 GB | 36W |
| Qwen3.5-9B | IQ2_XS | 91.30 | 6.90 GB | 38W |
| Qwen3.5-27B | IQ1_S | 42.10 | 24.50 GB | 39W |
Jetson running 27B IQ1_S = 42 tok/s @ 39W — half the speed of MBP 13 M1 (90 tok/s on 27B IQ2_XS), but 1/4 the power. Per joule: Jetson 27B = 1.08 tok/J vs MBP M1 27B = 0.27 tok/J (Jetson 4× better).
Edge device build recommendation
| Scenario | Recommended device | Model | tok/s | Power |
|---|---|---|---|---|
| Robot / drone | Jetson AGX Orin 64GB | Qwen3.5-9B IQ2_XS | 91 | 38W |
| Phone local LLM | Snapdragon 7 Gen 4 | Qwen3.5-0.8B Q4_K_M | 78 | 21W |
| Auto LLM | Snapdragon X Elite | Qwen3.5-4B Q4_K_M | 80 | 22W |
| Industrial edge server | Jetson AGX Orin | Qwen3.5-27B IQ1_S | 42 | 39W |
Key: all these devices can run LLM on battery power — this is the real value of edge AI.
Jetson vs Snapdragon
| Dimension | Jetson AGX Orin | Snapdragon X Elite |
|---|---|---|
| Avg tok/s | 110.96 | 80.82 |
| Avg power | 37.5W | 22.5W |
| Per-watt tok/s | 2.96 | 3.59 |
| VRAM | 64GB unified | 16GB unified |
| Supported models | 27B IQ1_S | 4B Q4_K_M |
| Best for | Industrial / auto / robotics | Phone / laptop |
Jetson is 37% faster but 17% less efficient per watt — Snapdragon is "power first", Jetson is "performance first".
Evidence
- 110.96/80.82/78.59 averages: DB 108 records
- 2.96 / 3.59 / 3.71 per-watt: DB actual power_w field
- 27B 4× tok/J: based on power comparison
- "Snapdragon 7 Gen 4 phone chip": Qualcomm official spec
Sources
- 45 Jetson Orin records: DB
benchmark_recordtable, hardware_id=76 - 36 Snapdragon X Elite records: DB benchmark_record, hardware_id=104
- 27 Snapdragon 7 Gen 4 records: DB benchmark_record, hardware_id=105