Domestic GPU Begins to Build Worlds: China's First Full-Stack Embodied Intelligence Simulation Platform Arrives

Moore Threads, a leading Chinese GPU manufacturer, has launched MT Lambda, the country's first full-stack domestically developed simulation platform for embodied intelligence. The announcement came during a recent event where the company demonstrated a remarkable feat: a robot dog trained entirely in a virtual environment performed complex maneuvers in the real world with perfect fidelity.

The robot dog, named Xiaofei, walked onto the stage, then performed a side flip in the simulation environment displayed on a screen. Instantly, the physical robot executed the identical motion. This sim-to-real transfer was achieved without any loss of performance, showcasing the platform's ability to train policies entirely in simulation and deploy them directly onto physical hardware.

MT Lambda is structured as a pipeline for physical AI training. At its core are two primary platforms: MT Lambda-Lab, focusing on strategy development and training for reinforcement learning, imitation learning, and VLA models; and MT Lambda-Sim, which provides high-fidelity physics simulation and rendering for scene construction, sensor simulation, data generation, and validation. Together, they form a closed loop: data synthesis → strategy training → simulation validation → edge deployment.

The platform is underpinned by three engines. The physics engine integrates multiple backends including MuJoCo-Warp-MUSA and Moore Threads' proprietary AlphaCore, achieving up to 30x simulation throughput under typical loads. The rendering engine, MT Photon, combines ray tracing and hybrid rendering, along with 3DGS and AI-generated rendering, to produce photorealistic visuals critical for accurate sensor simulation. The AI engine, deeply integrated with PyTorch via Torch-MUSA, supports VLA model development and reinforcement learning.

A key differentiator is Moore Threads' full-function GPU approach. Unlike specialized TPUs or NPUs, full-function GPUs simultaneously handle AI compute, graphics rendering, physical simulation, scientific computing, and video codec on a single chip. This is essential for embodied intelligence, which requires diverse workloads. MT Lambda leverages the MUSA architecture to unify these capabilities, eliminating the need to juggle multiple hardware and software stacks.

Moore Threads also outlined a broader ecosystem. On the cloud side, the KUAE cluster, powered by MTT S5000 GPUs (offering up to 1000 TFLOPS and hardware ray tracing), serves as a massive training ground. On the edge, the Changjiang SoC and E300 AI module provide 50 TOPS of local compute for real-time response on robots. Partnerships with companies like Zhiyuan, Guanglun Intelligence, and Pony.ai demonstrate real-world validation, from training large models to generating high-confidence simulation data.

The significance of MT Lambda extends beyond a single product. It represents a shift for domestic GPU makers from competing on chip specifications to building infrastructure for physical AI. By creating a platform that efficiently generates realistic, controllable, and scalable training worlds, Moore Threads aims to lower the barriers for embodied intelligence development. As the company's CEO stated, the goal is to empower all intelligent agents, from digital to physical, with a complete stack from cloud training to edge execution.

While challenges remain—such as simulation fidelity and ecosystem maturity—the launch marks a pivotal step. For the first time, a domestic GPU company has assembled a full-stack solution that can train, simulate, and deploy embodied intelligence entirely on homegrown hardware. This could accelerate progress in robotics, autonomous driving, and other physical AI applications, ultimately bringing intelligent machines closer to real-world deployment.