XPENG Unveils The “World Model Accelerator” X-Cache, Which Requires No Training, Is Plug-And-Play, And Boosts Inference Speed By 2.7 Times

XPENG has introduced a significant advancement in autonomous driving technology with the unveiling of X-Cache, a novel world model accelerator that boosts inference speed by up to 2.7 times without requiring any retraining. This plug-and-play solution leverages the continuity inherent in driving footage to identify and reuse image regions across consecutive video segments, thereby reducing redundant computations. By partitioning video data into temporally continuous segments and comparing intermediate features, X-Cache can skip entire layers of computation when minimal variation is detected, significantly enhancing processing efficiency while maintaining safety and visual fidelity. The innovation behind X-Cache lies in its ability to exploit physical continuity in the environment, recognising that many elements such as road surfaces and roadside features remain largely unchanged between frames. To ensure accuracy and prevent error accumulation, X-Cache generates a “fingerprint” combining driving actions and visual structures, and incorporates a safety mechanism that triggers full computation during critical scene transitions like turns or lane changes. This approach allows the accelerator to maintain high-quality output and system stability, making it suitable for real-world deployment in complex urban and highway scenarios. X-Cache has already been integrated into XPENG’s autonomous driving world model, X-World, demonstrating robust performance and engineering reliability. The technology achieves a 71% block skip rate, translating to a 2.6 to 2.7 times speed increase in inference with virtually no degradation in visual quality. This efficiency gain addresses one of the main bottlenecks in autoregressive video diffusion models, enabling high-frequency, large-scale simulation and interaction essential for the continuous evolution of autonomous driving systems. Beyond immediate performance improvements, X-Cache represents a strategic step towards scalable, cost-effective deployment of autonomous driving simulations. It forms part of a broader architecture in which XPENG’s VLA 2.0 handles perception and decision-making, X-World supports virtual-real mapping and scenario inference, and X-Cache accelerates inference processes. Together, these components enable a closed-loop system that integrates data acquisition, model training, simulation verification, and iterative development, pushing autonomous driving closer to full-stack, model-driven optimisation. This breakthrough in compute infrastructure not only enhances XPENG’s technological capabilities but also sets a foundation for expanding autonomous driving ecosystems with scalable, high-concurrency simulation. By efficiently utilising simulated worlds, XPENG is poised to reduce costs and improve operational capacity, marking a crucial evolution from simply building high-quality models to deploying them effectively at scale.