Newshot

No game engine needed! Musk says Tesla FSD v15 Reality View uses only half the compute of one TRIP accelerator on the AI4 chip

AI summary

Tesla has launched FSD v15 in the US market, introducing a new visualization interface called "Tesla Reality View." Musk said the feature is generated by real-time AI video without a game engine, using only half the compute of one TRIP accelerator on the AI4 chip.

Translation(model translation, not the source's own text)

No Game Engine Needed! Musk Says Tesla FSD v15 Reality View Uses Only Half the Compute of One TRIP Accelerator on the AI4 Chip

IT Home, November 22 news: Tesla today announced the launch of FSD v15 in the US market, bringing a brand-new visualization interface — "Tesla Reality View."

The new "Tesla Reality View" quickly sparked heated discussion among netizens. Blogger @旌旗超智能 analyzed that Tesla performed full-scene 3D Gaussian reconstruction despite a hardware disadvantage. He said he carefully examined the specs of Tesla's second-generation FSD chip today and found it inferior to the in-house chips of many new automakers.

IT Home noted that Musk also posted today explaining the feature. He revealed that FSD v15's "Tesla Reality View" is generated by real-time AI video, with no game engine needed! And it uses only half the compute of one TRIP accelerator on the AI4 chip.

  • "Tesla FSD v15's brand-new visualization UI sparks discussion; sources say full-scene 3D Gaussian reconstruction was done despite a hardware disadvantage"
  • "Tesla launches FSD v15 overseas: safety-critical scenario handling improved 25x, brand-new visualization UI"
Further reading: According to previous disclosures by Ashok Elluswamy, Tesla's head of AI, the neural network parameter count of FSD v15 has expanded from about 1 billion to about 10 billion, with a tighter link between prediction and control. Tesla positions v15 as a performance leap on par with V13 to V14, and says the upgraded vision encoder is shared across FSD, Robotaxi, and the Optimus humanoid robot, trained on millions of internet videos. FSD v15 runs the same software as Tesla's Robotaxi fleet. Elluswamy disclosed at the Q2 2026 earnings call that Robotaxis running early v15 versions had driven more than 380,000 miles unsupervised across six cities in two states, with no major accidents during that period. Regarding "Tesla Reality View," Tesla's release notes describe it as generating, in real time via the in-car AI computer, a surrounding road environment view from a virtual perspective above the vehicle. The interface supports full-screen, map, or split-screen views, and users can even tap a target location on the visualization to direct the vehicle there; Musk described it as "re-presenting real roads like a video game." The AI4 referred to in this article is Tesla's fourth-generation autonomous driving hardware platform, HW4. Since FSD V13, Tesla has renamed the HW4 chip platform AI4; the platform is built for FSD V13 and later versions, with about 5x the compute of HW3. According to previously leaked specs, HW4's CPU cores increased from 12 to 20, with a maximum frequency of 2.35GHz; its TRIP deep neural network accelerator upgraded from dual-core to tri-core, with maximum frequency raised from 2GHz to 2.2GHz, for an estimated total compute of about 300-500 TOPS. The "3D Gaussian reconstruction" mentioned in this article refers to 3D Gaussian Splatting (3DGS), a radiance field reconstruction method that explicitly represents a scene with a large number of 3D Gaussian ellipsoids and then renders novel views in real time via rasterization. The scene is explicitly represented by many ellipsoidal Gaussian particles, which are projected into 2D images to synthesize pictures, capable of generating novel views at 1080p resolution at over 30 frames per second. During training, the particle count starts small, and positions, shapes, and colors are adjusted by continuously comparing against real photos, with particles duplicated or deleted; during rendering, visible particles are spread onto the image like small color patches, and since they exist only where there is actual geometry, compute can be focused on the visible parts. Compared with neural radiance fields, which require multi-layer neural networks to query color and density point by point, its rasterization process is faster on graphics processors.

Further Reading

According to previous disclosures by Ashok Elluswamy, Tesla's head of AI, the neural network parameter count of FSD v15 has expanded from about 1 billion to about 10 billion, with a tighter link between prediction and control. Tesla positions v15 as a performance leap on par with V13 to V14, and says the upgraded vision encoder is shared across FSD, Robotaxi, and the Optimus humanoid robot, trained on millions of internet videos.

FSD v15 runs the same software as Tesla's Robotaxi fleet. Elluswamy disclosed at the Q2 2026 earnings call that Robotaxis running early v15 versions had driven more than 380,000 miles unsupervised across six cities in two states, with no major accidents during that period.

Regarding "Tesla Reality View," Tesla's release notes describe it as generating, in real time via the in-car AI computer, a surrounding road environment view from a virtual perspective above the vehicle. The interface supports full-screen, map, or split-screen views, and users can even tap a target location on the visualization to direct the vehicle there; Musk described it as "re-presenting real roads like a video game."

The AI4 referred to in this article is Tesla's fourth-generation autonomous driving hardware platform, HW4. Since FSD V13, Tesla has renamed the HW4 chip platform AI4; the platform is built for FSD V13 and later versions, with about 5x the compute of HW3. According to previously leaked specs, HW4's CPU cores increased from 12 to 20, with a maximum frequency of 2.35GHz; its TRIP deep neural network accelerator upgraded from dual-core to tri-core, with maximum frequency raised from 2GHz to 2.2GHz, for an estimated total compute of about 300-500 TOPS.

The "3D Gaussian reconstruction" mentioned in this article refers to 3D Gaussian Splatting (3DGS), a radiance field reconstruction method that explicitly represents a scene with a large number of 3D Gaussian ellipsoids and then renders novel views in real time via rasterization. The scene is explicitly represented by many ellipsoidal Gaussian particles, which are projected into 2D images to synthesize pictures, capable of generating novel views at 1080p resolution at over 30 frames per second. During training, the particle count starts small, and positions, shapes, and colors are adjusted by continuously comparing against real photos, with particles duplicated or deleted; during rendering, visible particles are spread onto the image like small color patches, and since they exist only where there is actual geometry, compute can be focused on the visible parts. Compared with neural radiance fields, which require multi-layer neural networks to query color and density point by point, its rasterization process is faster on graphics processors.

Advertisement statement: External redirect links contained in this article (including but not limited to hyperlinks, QR codes, passcodes, etc.) are used to convey more information and save selection time; the results are for reference only. All articles on IT Home containing external links include this statement.

Related Articles

  • Tesla FSD v15's brand-new visualization UI sparks discussion; sources say full-scene 3D Gaussian reconstruction was done despite a hardware disadvantage
  • Tesla launches FSD v15 overseas: safety-critical scenario handling improved 25x, brand-new visualization UI
  • US company heavily converts a Tesla Model 3 into an MPV, with supercharging, OTA, and driver assistance all usable
  • Dario rejected Musk's recruitment in 2010, joined OpenAI in 2016, then left to found Anthropic
  • The only one in the world: Musk visits Shanghai factory and receives a special "scrap-material figurine" from employees; Tao Lin posts photos
  • Already approved in 12 countries in Europe, Tesla's driver assistance system receives operating permit in Italy

The translation is machine-generated and individual terms may be off. Nothing is rewritten or cut; the original is authoritative.

Published by IT之家
This site only summarises and aggregates; copyright stays with the original author. Headlines and summaries follow the source — where they differ, the original wins.
Read the original ↗