VUDA Breaks CUDA-Vulkan Isolation for GPU Spatial Sharing
A team of researchers has introduced VUDA, a novel system that resolves the execution isolation issue between CUDA compute tasks and Vulkan graphics operations on a single GPU. This advancement facilitates spatial parallelism, permitting simultaneous execution of simulation and rendering processes in embodied AI settings. The study tackles a significant constraint where CUDA and Vulkan contexts are assigned to separate scheduling groups, necessitating exclusive time slices. Current spatial-sharing methods are restricted to CUDA, while temporal-sharing fails to optimize resources effectively. VUDA is designed for applications such as simulation data generation and reinforcement learning training, enhancing GPU utilization via spatial multiplexing. The research paper can be found on arXiv with the ID 2605.01352.
Key facts
- VUDA breaks execution isolation between CUDA and Vulkan on the same GPU.
- It enables spatial parallelism for compute and graphics workloads.
- Embodied AI environments interleave physics simulation (CUDA) and rendering (Vulkan).
- Simulation data generation and RL training can execute phases concurrently.
- Existing spatial-sharing techniques are limited to the CUDA ecosystem.
- Temporal-sharing approaches underutilize available resources.
- CUDA and Vulkan create separate GPU contexts with different scheduling groups.
- The paper is published on arXiv with ID 2605.01352.
Entities
Institutions
- arXiv