How Baidu’s Dual‑Engine GPU Container Virtualization Boosts AI, Rendering, and Cloud Gaming

1. Dual‑Engine GPU Container Virtualization 2.0

Last year Baidu released the industry’s first dual‑engine GPU container virtualization architecture, offering both user‑mode and kernel‑mode engines to meet diverse isolation, performance, and efficiency requirements.

On top of the isolation engine lies a resource‑pooling layer that decouples and pools resources via remote calls. Above that is a unified Kubernetes scheduling layer that supports various mixing strategies such as shared, preemptive, time‑slice, and tidal mixing, enabling AI model development, training, and online inference with higher GPU utilization.

Version 2.0 adds isolation for GPU rendering compute and codecs, achieving unified scheduling of AI, rendering, and codec resources.

2. New Capability Technical Analysis

The AI compute path uses CUDA, while rendering uses OpenGL/Vulkan; both share the same GPU compute resources. Although initial analysis suggested rendering could run within the AI isolation environment, experiments showed differences in command sets, leading to kernel‑mode implementation of rendering isolation.

User‑mode isolation is difficult due to the need to intercept many library calls and lack of transparency, so kernel‑mode isolation was chosen for rendering and codec workloads.

Kernel‑mode codec instances share weight with AI and rendering workloads, allowing fine‑grained allocation (e.g., 20% of codec power).

3. Full‑Scenario Practices

Different workloads benefit from different engines: online inference prefers user‑mode for low latency, while rendering‑heavy scenarios (e.g., autonomous‑driving simulation) require kernel‑mode isolation.

By using multi‑scheduler support within a single Kubernetes cluster, Baidu enables seamless coexistence of custom and Baidu schedulers, allowing label‑based GPU pool segregation and smooth migration.

In development environments, kernel‑mode virtualization provides multi‑user isolation and burst scheduling, improving GPU usage from ~20% to 35% and enabling shared GPU resources for tasks such as model debugging and data processing.

For cloud gaming, kernel‑mode memory isolation ensures QoS for rendering memory, while user‑mode is avoided due to latency constraints.

Overall, the dual‑engine approach unifies GPU resource management, maximizes utilization, and supports a wide range of AI, rendering, and codec workloads across scenarios like recommendation services, autonomous‑driving simulation, and ARM‑based cloud gaming.