Tesla Dojo v2 (2025 Speculative)
:::warning Speculative Content
Specs on this page are based on Tesla's 2024-10 We Robot event + Elon Musk public statements + industry analyst projections. Tesla has not officially released complete Dojo v2 specs. Official data subject to actual 2025 H2 announcement.
:::
Overview
Tesla Dojo v2 (codenamed Dojo 2) is Tesla's second-generation training AI chip, expected to launch 2025 H2. Based on TSMC 3nm process, 1 PFLOPS FP8 dense compute, Tesla custom D2 chip, 2nd Gen Fabric interconnect. Paired with PyTorch + Tesla Dojo SDK, targeting FSD (Full Self-Driving) v13/v14 + Optimus robot + Robotaxi training.
Strategic significance: Tesla is the only company that both builds cars and develops its own AI training chips (Apple is mobile, Meta uses GPU clusters, Microsoft/OpenAI use NVIDIA). Dojo v2 is Tesla's key product for transitioning from NVIDIA H100 dependency to in-house.
Core Specifications (Speculative)
| Item | Spec |
|---|
| Architecture | Tesla D2 |
| Process | TSMC 3nm (N3) |
| Compute Cores | 256 SCDs per D2 chip (Single Core for Dojo) |
| SCD Compute | 4 TFLOPS FP8 dense (speculative) |
| D2 Chip | 1 PFLOPS FP8 dense (256 x 4 TF) |
| Tile | 5x5 D2 = 25 chips / tile |
| Tile Compute | 25 PFLOPS |
| Cabinet | 2 tiles = 50 chips / cabinet |
| Cabinet Compute | 50 PFLOPS |
| ExaPod | 8 cabinets = 400 chips / ExaPod |
| ExaPod Compute | 400 PFLOPS FP8 |
| Interconnect | 2nd Gen Fabric (2 TB/s bidirectional, vs D1 1.6 TB/s) |
| TDP (ExaPod) | ~150 kW |
| Mass Production | 2025 H2 |
| Price (ExaPod) | ~$5-10M (speculative) |
Tesla Dojo Evolution
| Generation | Launch | Process | Compute | Interconnect | Use Case |
|---|
| Dojo v1 | 2023-07 | 7nm | 362 PFLOPS ExaPod | 1.6 TB/s | FSD v12 |
| Dojo v2 | 2025 H2 | 3nm | 400-500 PFLOPS ExaPod | 2 TB/s | FSD v13/v14 |
| Dojo v3 (speculative) | 2027+ | 2nm | 1 EFLOPS ExaPod | 4 TB/s | Optimus Gen 3 / Robotaxi |
Dojo v1 (D1 Chip) Known Specs
| Item | Spec |
|---|
| Architecture | Tesla D1 |
| Process | TSMC 7nm |
| Transistors | 50 billion |
| Cores | 354 SCDs (Single Core for Dojo) |
| Per SCD | 64-bit scalar + 64x64 matrix + 16x16 SIMD |
| Per SCD Compute | 1 TFLOPS BF16 |
| D1 Chip | 362 TFLOPS BF16 |
| Tile | 5x5 = 25 D1 chips |
| Tile Compute | 9 PFLOPS BF16 |
| Cabinet | 2 tiles = 50 D1 chips |
| ExaPod | 10 cabinets = 500 D1 chips = 1.1 EFLOPS BF16 |
Dojo v1 ExaPod known: 1.1 EFLOPS BF16 (7 ExaPods deployed, ~7.7 EFLOPS total)
Dojo v1 vs Dojo v2 Comparison
| Metric | Dojo v1 (2023-07) | Dojo v2 (2025 H2 speculative) | Improvement |
|---|
| Chip | D1 | D2 | New gen |
| Process | 7nm | 3nm | New gen |
| SCD Count | 354 | 256 (speculative) | -27% |
| Per-SCD Compute | 1 TF BF16 | 4 TF FP8 (speculative) | 4x |
| Per-Chip Compute | 362 TF BF16 | 1 PF FP8 (speculative) | 2.7x |
| Tile | 5x5 = 25 chips | 5x5 = 25 chips | Same |
| Cabinet | 2 tiles = 50 chips | 2 tiles = 50 chips | Same |
| ExaPod | 10 cabinets = 500 chips | 8 cabinets = 400 chips | Optimized |
| ExaPod Compute | 1.1 EF BF16 | 400-500 PF FP8 | 30-50% improvement |
| Fabric Bandwidth | 1.6 TB/s | 2 TB/s | +25% |
| TDP | 1+ MW | ~150 kW | Major power savings |
Dojo SDK Software Stack
| Layer | Tool | Notes |
|---|
| AI Framework | PyTorch | Tesla primarily uses PyTorch (vs NVIDIA CUDA) |
| JAX | Experimental |
| Compiler | Dojo Compiler | Auto vectorization + matrix optimization |
| Runtime | Dojo Runtime | ExaPod scheduling |
| Python API | dojo.torch | Similar to torch.device('dojo') |
| Distributed | Dojo Fabric | Cross-ExaPod communication |
| Visualization | Dojo Visualizer | ExaPod 3D real-time monitoring |
Dojo software advantage: PyTorch native support (vs Cerebras requiring SDK conversion), Tesla in-house 5 years.
Tesla In-House Dojo Strategy
| Dimension | 2023-2024 | 2025-2026 Speculative | 2027+ Speculative |
|---|
| FSD Training | Dojo v1 + NVIDIA H100 | Dojo v2 | Dojo v3 |
| Robots | - | Optimus training | Optimus Gen 3 |
| Robotaxi | - | Robotaxi training | Commercialization |
| Dojo Deployment | 7 ExaPods | 20+ ExaPods | 50+ ExaPods |
| Compute | ~7.7 EF | ~10-20 EF | 50+ EF |
| NVIDIA Dependency | 50% training | 30% training | 10% training |
| Cost Savings | - | ~$2B/yr vs all NVIDIA | $5B/yr |
Dojo strategic significance: Tesla is the only company that develops its own training chips + inference chips (FSD Computer) + data (Tesla Fleet). Dojo enables Tesla to fully vertically integrate AI.
Customers (Tesla Internal Only)
| Scenario | Use |
|---|
| FSD v12 (2024-Q4) | End-to-end neural network training |
| FSD v13 (2025 H1) | Dojo v1 training |
| FSD v14 (2025 H2) | Dojo v2 training |
| Optimus Gen 2 (2025) | Robot AI training |
| Robotaxi (2026) | L4-L5 autonomous driving training |
| xAI Grok | Backup (Tesla xAI collaboration) |
| Item | Details |
|---|
| Company | Tesla, Inc. |
| Business Unit | Tesla AI / Dojo team (Palo Alto + Austin) |
| Dojo Team | 200+ engineers (ex-AMD / Apple / Intel) |
| Fab | TSMC 3nm (Dojo v2) |
| 2024 Investment | $5B+ (Dojo R&D + manufacturing) |
| Goal | FSD / Optimus / Robotaxi full-stack AI training |
| Status | Continuous iteration (annual new gen) |
Use Cases
- ✅ FSD autonomous driving (v13/v14 training)
- ✅ Optimus robot (Gen 2/3 training)
- ✅ Robotaxi L4-L5 (2026 commercialization)
- ✅ Tesla internal distributed training (20+ ExaPods)
- ✅ Cost savings (vs NVIDIA H100 $30K/card)
- ❌ External sales (Tesla internal only)
- ❌ CUDA compatibility (requires Dojo SDK migration)
- ❌ Inference deployment (training only)
vs NVIDIA H100 Cluster (Dojo v2 ExaPod vs 1,000x H100)
| Metric | Dojo v2 ExaPod (400 chips) | 1,000x NVIDIA H100 |
|---|
| FP8 | 400 PF | 1.5 PF sparse (400W/H100) |
| FP16 | ~200 PF | 1 PF sparse |
| TDP | 150 kW | 700 kW |
| Efficiency | 2.67 TF/W | 2.16 TF/W |
| Price | ~$5-10M | ~$25-30M |
| Software | Dojo SDK | CUDA |
| Train LLM 405B | ~7 days (speculative) | ~3 days |
| Train FSD Network | Hours | Days (Tesla reported) |
Dojo vs NVIDIA cluster: 1.2x efficiency + 50% price + FSD training optimized, Tesla 2025+ reduces NVIDIA dependency by 50%.
Dojo v2 Key Timeline (Speculative)
| Date | Event |
|---|
| 2023-07 | Dojo v1 ExaPod activated (Tesla 7 ExaPods) |
| 2024-08 | FSD v12 end-to-end NN (partial Dojo training) |
| 2024-10 | We Robot unveils Optimus Gen 2 / Robotaxi (mentions Dojo v2) |
| 2025 H2 | Dojo v2 launch (speculative) |
| 2026 | Optimus Gen 3 + Robotaxi commercialization (Dojo v2 training) |
| 2027+ | Dojo v3 launch (2nm) |
Key Features
- D2 custom chip: Tesla 100% in-house (vs NVIDIA commercial)
- 3nm TSMC: Same generation as NVIDIA B200
- 1 PFLOPS per chip: Industry-leading training-dedicated
- 2 TB/s Fabric: High bandwidth across ExaPods
- PyTorch native: vs Cerebras requiring SDK conversion
- FSD / Optimus / Robotaxi full stack: Tesla unique vertical integration
- Weaknesses: Tesla internal only, 2-year ecosystem, single Tesla customer