Product Overview
Tesla Dojo is Tesla's custom-built AI training supercomputer. The D1 chip is purpose-built for Full Self-Driving (FSD) video data training. With 50 billion transistors, BF16 compute of 36.7 TFLOPS, 6 D1 chips form a Tile, a Cabinet integrates 6 Tiles, and an ExaPOD integrates 10 Cabinets (360 D1 chips).
D1 Chip Core Specifications
| Item | Spec |
|---|
| Architecture | Tesla D1 |
| Process | TSMC 7nm |
| Transistor Count | 50 billion |
| Compute Cores | 354 |
| FP32 | 9.1 TFLOPS |
| BF16/FP16 | 36.7 TFLOPS |
| INT8 | 36.7 TOPS |
| TDP | 400 W |
| Interconnect | Dojo proprietary interconnect (400W interface) |
Dojo System Topology
| Level | Configuration |
|---|
| D1 Chip | 1× 50 billion transistors |
| Tile | 6× D1 |
| Cabinet | 6× Tile |
| ExaPOD | 10× Cabinet = 360 D1 chips |
| ExaPOD Compute | ~100 EFLOPs (FP16 sparse) |
Manufacturer Info
Key Features
- Tesla full-stack in-house: chip + system + software
- Purpose-built for FSD video training: bypasses 4D annotation bottleneck
- ExaPOD 100 EFLOPs: single-cluster compute among the highest globally
- Liquid cooling
Use Cases
- Tesla FSD autonomous driving model training
- Large-scale video data processing
- End-to-end autonomous driving