New AI Chip and 3D Scene Model Target Generation Bottlenecks

According to tech website Leiphone, SmarCo GC3 is built to fight the 'memory wall' in video generation. Large diffusion-based video models constantly move data between compute units and off-chip memory, making data transfer the primary source of energy and latency rather than actual calculation. The chip abandons the program-counter and out-of-order execution model seen in conventional GPUs. Its dataflow architecture allows every computing node to start as soon as all needed inputs are ready, with the result flowing directly to the next node. Intermediate features stay on-chip until a full frame is generated, reducing repeated reads and writes and avoiding global sync waits. Leiphone notes the approach is structurally aligned with iterative, time-dependent diffusion backbones, although true production performance will depend on compiler quality and the maturity of the RISC-V AI ecosystem.

The chip carries 200 TOPS of INT8 AI compute, a 128 GB LPDDR5 unified memory pool with ECC, and 128 1080p hard-decode engines, allowing decoding, preprocessing, generation and post-processing in a single workflow. Zhongke Tongliang pairs dataflow with RISC-V vector extensions, turning high-frequency Transformer operators into dedicated execution units. It says this blend gives developers an efficiency ceiling from dataflow and an ecosystem floor from the open RISC-V community.

Quantum Bit reported that Hyper3D's WorldGen is the productized form of CAST, the scene-generation framework that took the best-paper award at SIGGRAPH 2025. Rather than creating the whole scene as one inseparable model, WorldGen identifies objects in the input image and generates each foreground object as an independent asset, while completing the background with 3D Gaussian splatting. In the default workflow a user uploads a picture and receives the scene in two to three minutes. In SimReady mode the system estimates collision geometry, mass, friction and restitution, so objects can be moved, swapped or deleted without breaking the environment.

Hyper3D founder and CEO Wu Di explained that old holistic pipelines merge every object into a monolithic block, making later edits and reuse difficult. CTO Zhang Qixuan defined WorldGen as a world generation model: it reproduces a scene from a text prompt or image into a usable three-dimensional world and acts as a carrier for broader world models. Hyper3D spent almost a year converting CAST into a product, mainly by reducing cascading errors from chained modules, according to Quantum Bit.

The tool is already being tested in production workflows. For games, WorldGen can turn a concept image into an editable scene draft so level designers can inspect paths and replace props before artists polish key assets; generated assets go into Blender, Unity, Unreal Engine and other DCC tools. Hyper3D forged an engineering partnership with Unity China in July. For embodied AI, Hyper3D, D-Robotics and simulation firm MouXianfei demonstrated a closed-loop pipeline in July: WorldGen generates interactive scenes from real images, MouXianfei's MotrixSim engine supplies physics and parallel simulation, and D-Robotics offers compute and developer toolchains. Filmmakers also combine WorldGen with video generator Seedance 2.5, using the 3D scene to stabilize camera movement and the video model to add performance, shading and visual polish.

Together, the two announcements point to where AI content generation is heading: removing the hardware and content-pipeline barriers that keep it out of professional use. SmarCo GC3 offers an alternative to stacking more GPUs for video generation, while WorldGen changes the base unit of 3D AI from a single object to an editable, simulation-ready scene. Both still need broad real-world verification, but they give designers and engineers a concrete path forward.