On the AGX Orin 32GB you get 3.5× the CUDA cores (1792 vs 512). The AI TOPS jump looks bigger than it is, because the two modules quote TOPS on different bases. The power envelope also moves up, from 10–30 W to 15–40 W.
Jetson AGX Xavier 32GB vs. Jetson AGX Orin 32GB developer... +3developer.nvidia.com ↗“512 CUDA + 64 Tensor cores, 8-core Carmel, 32GB LPDDR4x, 10-30W”nvidia.com ↗“up to 200 INT8 TOPS (sparse), 1792 CUDA + 56 Tensor cores”nvidia.com ↗“performance.max_ai_tops_super=241 TOPS”nvidia.com ↗“power.power_envelope_super=15-60 W”
| Spec | Jetson AGX Xavier 32GB | Jetson AGX Orin 32GB |
|---|---|---|
| GPU architecture | Volta | Ampere |
| CUDA cores | 512 | 1792 |
| AI performance (INT8) | 32 TOPS (dense) | 200 TOPS (sparse), default mode |
| AI performance, Super mode | No Super figure listed | 241 TOPS |
| Power envelope | 10–30 W | 15–40 W, default mode |
| Power envelope, Super mode | No Super figure listed | 15–60 W |
Things to keep in mind for the AMR migration:
- TOPS figures aren't directly comparable. Xavier's 32 TOPS is dense INT8, but Orin's 200 TOPS is sparse INT8. Some of the headline difference comes from that change of basis, not just from the hardware.
- Power budget. Orin's lower limit is 15 W, compared with Xavier's 10 W. Its default upper limit is 40 W, and up to 60 W with the Super figures. Check your battery and thermal budget against the mode you plan to run.
- Same form factor. Both modules use a 699-pin B2B connector and measure 100x87 mm. That doesn't mean your carrier board will work with Orin unchanged, so check with your carrier vendor. developer... +1developer.nvidia.com ↗“mechanical.form_factor=699-pin B2B (100x87mm)”nvidia.com ↗“mechanical.form_factor=699-pin B2B (100x87mm)”