Key Points

• Tesla AI FSD v13 Update Highlights: Tesla has shipped its latest version to 50,000 customers, enhancing highway autonomy significantly.
• Version 13 (v13) Upgrade: v13 introduces major scaling and architecture improvements, including 3x model scaling, audio inputs for emergency vehicles, and a redesigned control system for smoother driving.
• Silicon Performance: AI4/HW4 silicon achieves 1.3 Giga Pixels per second of inference power, comparable to or exceeding NVIDIA A100 in key metrics, with near-zero latency.
• Cortex Supercomputer Expansion: Tesla deployed 29,000 NVIDIA H100 GPUs with a target of 100,000 GPUs, supported by a new cooling plant with up to 350 MW capability.
• Gigafactory Supercomputer Power: The supercomputer at Giga Texas will draw 130 MW initially, scaling to 500 MW, with enough liquid cooling capacity to support this level of power draw.

As October comes to a close, Tesla AI has provided a detailed update on their latest software releases and advancements with the Tesla AI FSD v13 Update. These developments highlight Tesla’s commitment to improving Full-Self Driving (FSD) capabilities, leveraging cutting-edge technology, and scaling AI infrastructure. Here’s a quick breakdown of the key accomplishments and we’re the company is headed before the end of the year.

Recent Accomplishments

• End-to-End Highway Driving Deployment: Tesla successfully shipped the latest version (v12.5.6.1) of its end-to-end highway driving system to approximately 50,000 customers. This release represents a significant milestone in deploying more intuitive autonomous driving experiences for Tesla owners.
• “We, Robot” Event: Tesla hosted the “We, Robot” event, during which 50 autonomous Tesla vehicles transported over 2,000 passengers safely. This achievement showcases Tesla’s growing expertise in operating large-scale autonomous fleets in real-world conditions.

Upcoming Releases and Enhancements

1. Full Rollout of End-to-End Highway Driving: Tesla is preparing to extend the end-to-end highway driving feature to all AI4 users by early next week. Enhancements will include improved stop smoothness, reduced annoyance from bad weather notifications, and other safety-related improvements. This update should provide a more refined and seamless experience for users.
2. City Driving Model Improvements: The company is also focusing on improving AI3 city driving models. The new v12.5.x versions will aim to enhance vehicle behavior in complex urban environments.
3. Actually Smart Summon Expansion: Tesla plans to release the “Actually Smart Summon” feature to new regions, including Europe, China, and others. This expansion will allow more users worldwide to experience improved summoning capabilities, enhancing convenience and usability.

Major Upgrades with Version 13 (v13)

Tesla is gearing up for the release of v13, which is a major architectural overhaul:

• 36 Hz Full-Resolution AI4 Video Inputs: Improved sensor resolution and processing frequency for better perception.
• Native AI4 Input and Enhanced Neural Networks: Updates to the neural network architecture and adoption of AI4 inputs will increase accuracy and responsiveness.
• Scaling Enhancements: The v13 release includes 3x scaling of model size and context length, 4.2x scaling of data inputs, and 5x scaling of training compute power, enabled by Tesla’s new Cortex training cluster.
• Enhanced Collision Avoidance and Navigation: Using better reward predictions for collision avoidance and smoother navigation, v13 aims to make autonomous driving significantly safer and more reliable.
• Audio Input for Emergency Vehicle Handling: New audio input integration will help Tesla vehicles respond more effectively to emergency vehicles.
• Redesigned Control System: The controller system has been redesigned for smoother, more accurate tracking, and integrated with unpark, reverse, and park capabilities, making vehicle operations more seamless. This will also allow for users to select where they want the car to end the ride, such as finding a spot in a parking spot, finding street parking, pulling up to a curb, etc.
• Improved Occlusion and Cleaning Mechanisms for Cameras: Better handling of camera occlusions will ensure more consistent visual input, particularly in challenging environments.

Tesla’s silicon advancements, particularly with AI4/HW4 enable inference at a rate of approximately 1.3 Giga Pixels per second, achieved in a highly power-efficient manner. The AI4/HW4 hardware stack has been engineered to process high-resolution inputs at 36 Hz, allowing Tesla vehicles to operate with a level of responsiveness and perception that rivals human drivers. This performance is comparable to or exceeds NVIDIA A100 GPUs in key aspects such as latency, where Tesla achieves near-zero latency between photon capture and inference.

Infrastructure Scaling and Compute Advancements

In addition to AI model enhancements, Tesla continues to make substantial investments in AI infrastructure:

• Cortex Supercomputer Training Cluster: Tesla’s new Cortex supercomputer training cluster, based at Giga Texas, is playing a critical role in scaling FSD’s training compute. The upcoming v13 release is the first to be trained entirely using the Cortex cluster, representing a fivefold increase in training compute scaling.
• NVIDIA H100 GPU Deployment: Tesla recently deployed 29,000 NVIDIA H100 GPUs at Giga Texas, which has significantly increased their AI training compute capacity by 75% in Q3. By the end of October, Tesla expects to expand this capacity to 50,000 H100 GPUs, with a future target of reaching 100,000 H100 equivalent capacity.
• Cooling Infrastructure Expansion: As Tesla’s Cortex data center continues to grow, an operational liquid cooling plant at Giga Texas becomes vital to sustain and efficiently manage the increased heat generated by the growing compute resources. The cooling plant is designed with enough liquid cooling capacity to handle up to 500 megawatts, ensuring the infrastructure can accommodate the immense thermal load as compute demands increase. Musk estimates the Gigafactory supercomputer will draw 130 megawatts on deployment, with growth up to 500 MW expected after Tesla’s proprietary AI hardware is also installed.

Conclusion

Tesla’s October updates reflect a push towards expanding Full-Self Driving capabilities, particularly with the v13 release set to bring significant model and architectural upgrades and enhancements. The substantial scaling in AI training compute capacity and the integration of new, advanced features in the Tesla AI FSD v13 Update demonstrate Tesla’s holistic approach to autonomous vehicle technology. With the upcoming full rollout of v13.3 aimed for Thanksgiving, Tesla aims to get one step closer to true autonomy. The upcoming infrastructure expansions, including GPU scaling and cooling facilities, will further enhance Tesla’s ability to innovate and deliver new features at an accelerated pace. As we’ve seen with all AI based on transformers to date, compute and data improve end result.

The developments from Tesla AI showcase their ongoing drive to innovate in autonomous driving, while also paving the way for the future of AI and EV integration. These advancements highlight Tesla’s dedication to making autonomous technology a part of everyday life, pushing the limits of what’s achievable, and reinforcing their leadership in the future of AI-powered transportation.

Join the newsletter and stay up to date