🧵 REPLYING TO @samwose (RimRunner):
┌─ "1. Breaking Free from Nvidia
Tesla currently relies on Nvidia’s A100 and H100 GPUs for training its massive video-based neural networks. While powerful, these chips are general-purpose and optimized for broader markets like LLMs and gaming.
Dojo 2, by contrast, is Tesla’s application-specific supercomputer, built from the ground up to train vision-centric, spatiotemporal neural nets for FSD, Optimus, and Grok.
With Dojo 2, Tesla moves from customer to competitor - gaining control over:
• Cost per FLOP
• Power efficiency
• Availability
• Scaling roadmap
2. TSMC’s INFO-SOW: Advanced Chip-on-Wafer Stacking
Tesla is using TSMC’s cutting-edge Integrated Fan-Out with Silicon-On-Wafer (INFO-SOW) packaging tech. This enables:
• High-bandwidth die-to-die connections
• Lower latency between training modules
• Improved heat dissipation for dense compute clusters
In massive AI workloads, memory bottlenecks and heat limits matter more than raw compute. Dojo 2’s architecture addresses both.
3. Tesla’s Custom Instruction Set & Training Stack
Unlike general-purpose GPUs, Dojo cores are designed to accelerate matrix multiplies, systolic arrays, and video-token fusion - the backbone of FSD’s vision-based planning stack.
By building its own silicon, Tesla can:
• Fuse model architecture and compiler logic
• Optimize for batch sizes and data locality
• Avoid CUDA-style abstraction penalties
4. Strategic Leverage
Dojo 2 gives Tesla:
• Internal cost control at exascale
• Scalable infrastructure for FSD rollouts
• A competitive moat in AI training
• Infrastructure for non-driving AI (Optimus, Grok agents, multimodal inference)"
└─────────────────────────────────
💬 ELON'S REPLY:
@samwose @StockSavvyShay It’s a good computer