Yandex's 3D reconstruction team has successfully implemented the execution of game engines, including Unity, inside YTsaurus containers to automate the rendering of 3DGS avatars. This solution allowed for scaling the generation of heavy visual content using standard orchestration tools.
What Happened
Yandex engineers configured a full graphics stack inside an isolated container environment. To achieve this, not only basic GPU devices were passed through to the system, but also specific nodes: /dev/nvidia-modeset, /dev/dri/card0, and /dev/dri/renderD128. Additionally, an X11 server with the necessary DDX driver (nvidia_drv.so) and dbus was deployed inside the container. On architectures such as NVIDIA A100, which lack RT cores and the NVENC hardware encoder, the team applied Compute shaders for ray tracing and software video encoding on the CPU.
Context
Traditionally, rendering tasks require specialized hardware with video outputs and desktop drivers, which makes scaling them in the cloud difficult. Using YTsaurus allows the process to transition from manual desktop session management or the use of heavy virtual machines into a format of managed microservices.
Why It Matters for the Industry
This case marks a transition from using GPUs exclusively for neural network training to a full-fledged infrastructure for generation and rendering (inference/generation). Transforming distributed computing clusters into scalable render farms allows the industry to process massive volumes of synthetic content, simulations, and 3DGS objects as efficiently as model training.
Why It Matters for Users
For developers, this opens the way to creating automated pipelines for mass content production—for example, creating thousands of videos with 3D avatars (similar to the "Cybervillage" project) within a single workflow without the need to manually launch engines.
Sources
Author
Look at AI, Editorial Team