gVulkan: Scalable GPU Pooling for Pixel-Grained Rendering in Ray Tracing
summary
gVulkan implements transparent API-forwarding for Vulkan to offload ray tracing to a cloud GPU pool, uses resource-classified API-forwarding to avoid round trips, skips local present work and decouples GPU/CPU stages for parallelism, and runs Split Frame Rendering with pixel-grained dynamic self-rebalancing to redistribute tiles across GPUs. Prototype shows 3.81× speedup on average with 4 GPUs, 10.3× higher FPS versus uncategorized forwarding, and returned images with PSNR >30, demonstrating scalable multi-GPU acceleration.
AI-written summary · about
chapters (9)
- 0:00gVulkan intro and ray tracing
- 2:54Real-time ray tracing limits
- 5:23Why Vulkan for gVulkan
- 7:29Multi-GPU ray tracing challenges
- 8:35How gVulkan solves these challenges
- 10:22gVulkan core components
- 11:46Three gVulkan methods
- 15:19Performance results
- 17:06Conclusion and future
related talks
01h 00mGPU Accelerated Computing & Optimizations on Cross-Vendor Graphics Cards with Vulkan & Kompute
25mVectorVisor: A Binary Translation Scheme for Throughput-Oriented GPU Acceleration
23mHello GPU
06mLightning Talk: Creating a GPU With C++ and an FPGA
31mGPUs Unleashed! Make Your Games More Powerful With wasi-gfx
31mWrite TypeGPU, Use Anywhere: Composable WebGPU Shaders for React Native and Beyond
25mPathfinder to GPU Offload in WASM
37mC++ for multi-accelerator, multi-vendor systems