PyTorch at Tesla
summary
Tesla uses PyTorch to train vision-first networks on raw video from eight cameras, running large multihead “HydraNets” with shared ResNet-50 backbones and a second recurrent topdown pass. Training totals 48 networks producing 1,000 output tensors and consumes about 70,000 GPU hours; inference targets the in-house FSD computer and future Dojo training hardware. Field results include 1 billion miles of Navigate on Autopilot, ~200,000 confirmed lane changes and ~800,000 Smart Summon sessions.
AI-written summary · about
chapters (13)
- 0:00Intro
- 0:47TESLA AUTOPILOT
- 1:06SMART SUMMON
- 1:24FULL SELF DRIVING
- 1:53FULL STACK AL
- 2:51SINGLE VIEW TASKS
- 4:30ACROSS-CAMERA TASKS
- 4:41RECURRENT TASKS
- 6:59MULTI-TASK TRAINING
- 7:29A FULL BUILD
- 8:55OPERATION VACATION
- 9:10FSD COMPUTER
- 10:25NAVIGATE ON AUTOPILOT
related talks
31mDeep learning for self-driving vehicle technology
19mAI Trends in Intelligent Transport Systems
42mAI and Self-Driving Cars
52mPhysical AI: Crafting Resilient Systems with Emotional Intelligence
40mFueling the AI Revolution with Gaming
20mAccelerating AI pipelines w/ NVIDIA on AWS: An autonomous driving case
47mUnlocking Data-Driven Apps with TensorFlow
40mMachines are Learning