Among Jetson production modules with a max power envelope of 20 W or less, the Jetson Orin NX 8GB offers the highest memory bandwidth: 102 GB/s (LPDDR5), within its default 10–20 W power envelope. nvidia.com +1nvidia.com ↗“memory.memory_bandwidth_gb_per_s=102”nvidia.com ↗“power.power_envelope=10-20 W”
For a memory-bound vision transformer workload, that makes it the clear pick in the default power mode. The full ranking of qualifying modules:
Memory bandwidth, ≤20 W max power envelope (default mode) nvidia.com +7nvidia.com ↗“memory.memory_bandwidth_gb_per_s=102”nvidia.com ↗“memory.memory_bandwidth_gb_per_s=68”developer.nvidia.com ↗“memory.memory_bandwidth_gb_per_s=59.7”developer.nvidia.com ↗“memory.memory_bandwidth_gb_per_s=59.7”developer.nvidia.com ↗“memory.memory_bandwidth_gb_per_s=59.7”developer.nvidia.com ↗“memory.memory_bandwidth_gb_per_s=51.2”nvidia.com ↗“memory.memory_bandwidth_gb_per_s=34”developer.nvidia.com ↗“memory.memory_bandwidth_gb_per_s=25.6”
| Module | Memory Bandwidth | Memory | Power Envelope |
|---|---|---|---|
| Jetson Orin NX 8GB | 102 GB/s | 8 GB LPDDR5 | 10–20 W |
| Jetson Orin Nano 8GB | 68 GB/s | 8 GB LPDDR5 | 7–15 W |
| Jetson TX2 | 59.7 GB/s | 8 GB LPDDR4 | 7.5–15 W |
| Jetson Xavier NX 16GB | 59.7 GB/s | 16 GB LPDDR4x | 10–20 W |
| Jetson Xavier NX 8GB | 59.7 GB/s | 8 GB LPDDR4x | 10–20 W |
| Jetson TX2i | 51.2 GB/s | 8 GB LPDDR4 | 10–20 W |
| Jetson Orin Nano 4GB | 34 GB/s | 4 GB LPDDR5 | 7–10 W |
| Jetson Nano | 25.6 GB/s | 4 GB LPDDR4 | 5–10 W |
A few notes relevant to your use case:
- Super mode: Jetson Orin NX 8GB and Orin Nano 8GB both also list a Super-mode memory bandwidth of 102 GB/s, but their Super-mode power envelopes extend to 40 W and 25 W respectively — so if you hold the strict 20 W ceiling, the default-mode figures above are the ones that apply. nvidia.com +3nvidia.com ↗“memory.memory_bandwidth_gb_per_s_super=102”nvidia.com ↗“power.power_envelope_super=10-40 W”nvidia.com ↗“memory.memory_bandwidth_gb_per_s_super=102”nvidia.com ↗“power.power_envelope_super=7-25 W”
- Compute headroom alongside bandwidth: the Orin NX 8GB delivers up to 70 INT8 TOPS (117 TOPS in Super mode) with an Ampere GPU and 1024 CUDA cores, which pairs well with transformer inference at this power point. nvidia.com +2nvidia.com ↗“performance.max_ai_tops=70 TOPS”nvidia.com ↗“performance.max_ai_tops_super=117 TOPS”nvidia.com ↗“1024 CUDA cores”
- If you can trade memory bandwidth for capacity, the Xavier NX 16GB doubles memory to 16 GB at 59.7 GB/s — but for a bandwidth-bound model, the Orin NX 8GB's 102 GB/s is the decisive advantage. developer... +1developer.nvidia.com ↗“16GB LPDDR4x”developer.nvidia.com ↗“memory.memory_bandwidth_gb_per_s=59.7”