ADAS refers to Advanced Driver Assistance Systems, a collection of terms and technologies that culminate in autonomous vehicles. SAE defines six levels of ADAS:
| SAE Level | Capabilities | Example |
|---|---|---|
| 0 | Alerts only | Lane departure alarm |
| 1 | Single system | Adaptive cruise control |
| 2 | Multiple systems | Adaptive cruise + lane centering |
| 3 | Conditional automation | Full driving in specific scenarios (not sure) |
| 4 | High automation | No driver input needed within geofenced areas. Waymo is Level 4. |
| 5 | Full automation | No driver under any conditions (weather, no geofence, etc) |
HPC requirements
The FalconFS paper describes one aspect of training models for ADAS very well.1 Its storage requirements are stated as:
- Multimodal data stored as one file per mode over time
- Images, point clouds
- “few KiB to a few MiB, mostly within 256 KiB”
- Inflight dataset can be up to hundreds of petabytes, 300 billion files
- Billions of directories (timestamps, vehicle ID, camera ID, …)
- Random access order of files
The CPUs are used for data augmentation in the loop, leaving little resources for ancillary services on compute nodes.