Replies

Cont’d

Key Purposes and Capabilities
Autonomous Driving Inference: These chips process real-time vision-based AI models for perception, planning, and control in Tesla vehicles, replacing or augmenting third-party GPUs (e.g., from NVIDIA). The A15/AI5, for instance, delivers up to 40x the performance of the prior AI4 (HW4) in key metrics, with 8x more raw compute power and 9x more memory, while running inference up to 10x cheaper than comparable NVIDIA chips.
Robotics and Broader AI Applications: Surplus production is allocated to power Optimus robot brains and edge computing in small data centers, focusing on low-power, high-efficiency designs (e.g., performance per watt and dollar optimized for scale).
Evolution Across Generations:
A15/AI5: Current focus; a “monster” chip (per Elon Musk) with a converged architecture that eliminates separate GPUs and image processors for a compact half-reticle design. Production starts in the US via Samsung (Texas) and TSMC (Arizona) in 2026.
A16/AI6: Follows closely, introducing dual capabilities for both inference and training— a key upgrade over prior chips—while maintaining the same multi-use versatility.
AI7: Planned for rapid succession after AI6; details are sparse, but it’s part of Tesla’s aggressive roadmap to push toward “out-of-this-world” AI8, emphasizing even greater scalability for real-time autonomy and robotics.
These chips stem from Tesla’s shift away from the Dojo training supercomputer team to prioritize inference-focused silicon, reducing reliance on external suppliers and accelerating FSD deployment.