In June 2026, the global top-tier computer vision conference CVPR opened in Denver, USA. Ashok Elluswamy, Vice President of AI Software at Tesla, officially disclosed at the conference that FSD V14 adopts a single end-to-end neural network with an inference frequency of 36 Hz, far exceeding the industry mainstream level of 10 Hz. XPeng was invited to the stage for the third time, comprehensively showcasing its world model technology map, which includes three core technologies: X-World, X-Foresight, and X-Cache, building a dual-pillar physical AI architecture of VLA + World Model. From these technical architectures, it is evident that the autonomous driving industry is undergoing a technological shift from software-defined to AI-defined.
01. What Problems Did Software-Defined Vehicles Solve? And What Problems Did They Leave Behind?
Software-defined vehicles primarily solved the problem of hardware-software decoupling. The functions of traditional vehicles were fixed at the factory and could not be changed later, whereas software-defined vehicles can continuously acquire new functions through OTA (Over-The-Air) upgrades, iterating continuously like smartphones. However, the limitations of software-defined vehicles are equally obvious: all its function iterations are confined within preset programs and fixed rule frameworks, belonging to passive, standardized function updates.
In the era of software-defined vehicles, engineers would write instructions into the code one by one, such as braking when a pedestrian is within ten meters or changing lanes when there is a cone ahead. This approach seems to patch system loopholes, but real-world driving scenarios are virtually endless, and engineers can never write all the rules. More importantly, software-defined solves the problem of being iterable, not the problem of being able to think. The difference between iterable and thinking is that one allows the vehicle to continuously install new functions, while the other allows the vehicle to judge for itself what to do.
When Tesla's FSD transitioned from V11 to V12, it underwent a critical shift. It replaced over 300,000 lines of C++ code in the V11 control stack used for urban street driving with an end-to-end neural network trained on millions of video clips. However, the V12 era did not achieve zero code; the system still retained about 2,000 lines of C++ code, mainly responsible for non-core control logic such as system scheduling. This transformation also proved that data-driven neural networks can surpass hand-coded rules in complex driving decisions, paving the way for subsequent thorough full-chain AI integration.
By V14.2 and earlier versions, even though the perception and planning modules had extensively adopted neural network technologies, the most core vehicle underlying control (including steering angle, throttle, and brake force adjustment) still relied on hand-written C++ code by engineers. This led to rigid vehicle driving performance, such as abrupt braking at yellow lights and repeated braking at stop signs. The reason for this issue is that code can only cover the scenarios engineers thought of in advance, and the parts outside the rules became the system's capability blind spots.
In April 2026, Tesla began rolling out FSD V14.3. This update removed all hand-written C++ code in the control link, replacing it with neural networks directly outputting driving instructions. From the system receiving images from onboard cameras to the final steering wheel and pedal instructions, there is no longer any constraint from hand-written rules in between. This is the full-domain unified end-to-end architecture. From mostly end-to-end in V12 to full-domain end-to-end in V14.3, it took Tesla two years to remove the last code from the control chain.
02. End-to-End: From Rule-Driven to Data-Driven?
The end-to-end neural network is the first technological breakthrough for AI-defined vehicles.
The so-called end-to-end means that through a single AI model, raw sensor data can be input to output final control instructions. The traditional modular architecture splits perception, planning, and control into independent modules, each using different algorithms, and information is passed between modules through manually defined interfaces; end-to-end uses a single neural network to cover the entire process from sensor input to control output. The essential difference between these two architectures is that modularization is about engineers writing rules to tell the system what to do, while end-to-end is about the system learning how to do it from data itself.
The biggest advantage of end-to-end is its generalization ability. Traditional rule systems are prone to errors in unseen new scenarios, whereas end-to-end models learn the overall mapping relationship from sensor input to driving decisions, enabling them to make relatively reasonable reactions in unfamiliar scenarios. The average takeover interval of FSD V14 in complex urban road conditions has been reduced to only 0.3 times per 1,000 miles, with safety approximately 8 times that of an average human driver in the US.
Tesla made several key technical improvements in V14.3. The model parameter size was increased to about 10 times that of previous versions, enabling it to remember more diverse road condition patterns, including edge cases such as irregular intersections, composite traffic lights, complex construction zone layouts, and non-standardized traffic signs. The system also added explicit spatiotemporal memory capabilities for the first time, with a duration of about 3 to 5 seconds. It is no longer a forgetful AI; it can remember who suddenly cut into the lane a few seconds ago and deduce the future 3 to 5-second trajectories of humans and vehicles based on this.
Tesla also made changes at the compiler level. Based on the MLIR (Multi-Level Intermediate Representation) framework, Tesla rewrote the AI compiler and runtime environment from scratch. It can retain the structural information of the computational graph at a higher level of abstraction, making it easier to identify which operations can be merged and which data can be reused, thus generating more streamlined hardware instruction sequences. The result of this change is that the time delay from the camera capturing the image to the vehicle executing the action is reduced by 20% compared to before. In an emergency evasion at a speed of 80 km/h, this means a braking distance of over one meter.
At CVPR 2026, Tesla further disclosed more technical details of FSD V14. The system introduced three "psychological scoring" dimensions: comfort score, intervention probability score, and human imitation discriminator score, enabling the vehicle to independently evaluate the anthropomorphic degree of the intended action. Tesla is also deeply applying world model technology, reconstructing real vehicle camera footage into 3D "skeleton" models through patents to generate nearly infinite training scenarios.
However, end-to-end is not without problems. A giant model directly ingesting sensor data and outputting driving instructions is like a black box; you don't know whether a certain hard brake was triggered by seeing a real danger or misinterpreting a shadow. Poor interpretability and difficult safety verification are also the core obstacles faced by regulatory authorities when approving L3 and above autonomous driving.
03. VLA and World Models: From Perception to Understanding
After end-to-end, the industry's technological revolution did not stop. VLA (Vision-Language-Action) models and world models have become new technological high grounds.
VLA introduces action tokens on the basis of vision-language models, allowing the model to jointly learn from the pairing of vision, language, and robot trajectory data. VLA integrates the three major modules of vision, language, and action. The model has a huge number of parameters, and the inference computation far exceeds traditional end-to-end algorithms.
Li Auto's Mach VLA system is a representative of the VLA route in China. This system is defined as an embodied intelligence system that fully connects perception, computation, decision-making, and execution. The Mach VLA large model is the brain of embodied intelligence, the 3D ViT perception model is the eyes, the Mach M100 chip is the heart, the Xinghuan OS is the nervous system, and the full wire-controlled chassis is the hands and feet.
As a pure vision solution, 3D ViT can output rich representations similar to full-color point clouds, including RGB information, texture features, and dynamic-static separation. The depth accuracy at close range is close to that of LiDAR. The multimodal computing performance of the Mach VLA 2.1 version is increased by ten times compared to the previous generation, deeply integrating LiDAR and visual data, effectively increasing the visual range by 50%. Through full-chain optimization, the system reduces the reaction time in emergency situations to 0.28 seconds, about 40% faster than an average human driver. In addition, the system's visual input latency is reduced by 47%, model inference latency is reduced by 43%, chassis response is reduced by 38%, and OS scheduling is reduced by 28%, with the overall end-to-end latency reduced by 40%. Li Auto plans to push a new version of Mach VLA to users in the third quarter of 2026, and in the fourth quarter, its overall intelligent driving capability will benchmark against Tesla's FSD V14.
XPeng took a different path, adopting a dual-pillar route of VLA and world models. Liu Xianming, head of XPeng's General Intelligence Center, explicitly responded to the long-debated route issue in the industry at CVPR 2026, stating that VLA and world models are not mutually competitive. In XPeng's architecture, the second-generation VLA learns how to act from human driving behavior, while the world model learns how the world will change after the action by predicting future states. The two together constitute the physical world foundation model. The former allows the model to learn how to act, and the latter allows the model to understand how the world will change after the action.
XPeng's second-generation VLA has billions of parameters, trained on over 100 million video clips. Each version of the model is trained with over 4 trillion tokens, and the on-vehicle inference speed is increased by 12 times. Real-world test data shows that VLA 2.0 can reduce hard braking instances by 99% and jerky movements by 89%, with complex road condition traffic efficiency on par with human drivers. In the first month of rollout, the proportion of user-assisted driving mileage exceeded 50%, making it the first system in the industry to achieve this data.
Regarding the world model, XPeng disclosed three core capabilities: active thinking, controllable generation, and long-sequence deduction. The corresponding technical achievements include X-Mind (active reasoning and decision-making interpretability), X-World (controllable multi-view generative world model), X-Foresight (vision-action causal prediction network), and X-Cache (inference acceleration solution).
X-World can generate future videos that comply with physical constraints given specific action conditions, maintaining good controllability and stability during continuous generation. The world model borrows the next-token prediction paradigm from large language models, gradually learning the dynamics and causal structure of the physical world through dense prediction of the next frame or next state on massive unlabeled videos. Relying on a 10,000-GPU intelligent computing cluster, XPeng increased its single-GPU training efficiency by 1010% in one year, and GPU hardware utilization increased from 40% to 90%.
Waymo also unveiled its own world model at CVPR 2026 for the first time. Waymo did not choose to train from scratch but directly used Google DeepMind's Genie 3 as the foundation model. As a world simulator that understands the operating rules of the real world, Genie 3 was added with Waymo's exclusive sensor data (multi-camera, LiDAR, millimeter-wave radar, high-precision maps) for mid-term training, and finally fine-tuned and distilled for specific driving tasks.
This model can predict the dynamic evolution of road scenarios, such as the behavioral intentions of other vehicles and pedestrians, as well as changes in traffic signals, endowing the system with predictive capabilities. The industry compares this direction to the Genie era of autonomous driving, where the system no longer passively perceives but actively understands and predicts the world. NVIDIA also released Cosmos 3, described as the world's first physical AI foundation model that unifies visual reasoning, world generation, and action generation.
From rule-based to end-to-end, and then to VLA and world models, the direction of technological evolution is becoming increasingly clear. Autonomous driving systems are moving from executing instructions to understanding the world.
04. Reconstruction of Computing Architecture: From TOPS to Bandwidth
In the era of AI-defined vehicles, completely different requirements are also posed to the underlying infrastructure.
A report released by McKinsey in June 2026 pointed out that autonomous driving is shifting from an automotive engineering problem to an AI infrastructure problem. The explosion in computing power demand brought by the end-to-end architecture has directly pushed chips from being a mere automotive component to a core variable determining competitiveness. In the global market for ADAS (Advanced Driver Assistance Systems) and autonomous driving processing semiconductors, it was about $7.6 billion in 2025 and will exceed $46 billion by 2035, with a compound annual growth rate of about 24%.
Looking back at the technology launch events in the past two years, when entering the era of large models, automakers have been competing on TOPS (Tera Operations Per Second). However, the real bottleneck in the end-to-end era is not computing power, but memory bandwidth. End-to-end models have a huge number of parameters, and sensor data streams are high-resolution, multimodal, and high-frame-rate. The volume of intermediate activation layer data far exceeds the model itself. Before the computing power is exhausted, the bandwidth is already blocked first. High-bandwidth LPDDR memory, larger on-chip SRAM caches, and even HBM are replacing TOPS numbers as the hard indicators for measuring the next-generation ADAS chips.
In addition, the chip architecture itself is also changing. The traditional Von Neumann architecture drives computation with sequential instructions, using a large number of transistors to handle management overheads such as caching, scheduling, and branch prediction. In AI computing tasks, the efficiency bottleneck of the traditional architecture will be further amplified.
The ideal solution is to adopt a dataflow architecture. AI computing is naturally parallel; data is in tensors, relationships are deterministic, and flow paths are clear. Data flow drives computation to occur, rather than being dominated by a central instruction queue.
Li Auto's self-developed Mach M100 Ultra adopts a 5nm automotive-grade process, with a single-chip computing power of 1280 TOPS. Inside the chip, Li Auto allocates more than half of the wafer area to the neural processing unit. The NPU part consists of 56 computing units and 1 data processing module. The memory system adopts an 8-channel LPDDR5X subsystem, and the off-chip memory bandwidth reaches 273 GB/s. More importantly, because of the dataflow architecture design, the actual computing power utilization rate exceeds 82%. During the operation of the end-to-end model, the system latency can be reduced by 40%, dropping to the 200 to 300 millisecond range.
BYD released the 4nm process intelligent driving chip Xuanji A3, equipped with a 3-core NPU, natively supporting Transformer large models, with a DDR bandwidth of 273 GB/s. Through the collaboration of three chips, it achieves a total computing power of over 2100 TOPS. The NPU is replacing the GPU as the main force for end-to-end chips, while the CPU is shifting to handle safety redundancy and system scheduling.
05. Final Words
In fact, at this stage, the direction of the evolution of autonomous driving technology is becoming increasingly clear. Vehicles are transforming from upgradable tools into thinking intelligent agents. Actually, this transformation has just begun, but its impact is rapidly spreading to chips, algorithms, data, and even the entire industrial landscape. Whoever can build real technological barriers in the AI-defined track will take the initiative in the next stage of competition.
#AutonomousDriving #SoftwareDefined #LargeModels