Editing Applications of Edge Computing (section)

===8.3.4 Computational Constraints===
Beyond basic AUV capabilities, Edge computing requires unique processing limitations.
* '''Real-Time Latency Bottlenecks'''
: {| class="wikitable" style="margin:auto"
|-
! Task !! Max Allowable Latency !! Typical Edge Performance
|-
| Obstacle avoidance || 100ms || 50–80ms (Jetson AGX Orin)
|-
| Fault detection || 500ms || 200–300ms (FPGA solutions)
|}

* '''Memory Limitations'''
: Most AUV edge systems have 4–16GB RAM vs. cloud servers' 100+GB. This is handled by a technique known as Tensor Slicing where large models are processed in chunks or by using MRAM caches which needs almost 10x lower power than SRAM for frequent data access.

These constraints drive innovation in edge solutions, which we explore in Section 8.4 through state-of-the-art architectures like federated learning and hybrid systems.