GPU-Based Visualization for Flight Simulation Tim Woodard Director of Research and Development Diamond Visionics www.dvcsim.com GPU Technology Conference 2014 Flight Simulation Takes Off Flight Simulation Takes Off Link’s Blue Box, 1943 Flight Simulation Takes Off Link’s Blue Box, 1943 Terrain model-board and film, 50’s and 60’s Flight Simulation Takes Off Link’s Blue Box, 1943 Terrain model-board and film, 50’s and 60’s “Image Generators”, 70’s to present “Big Iron” McDonnell Douglas VITAL Evans and Sutherland ESIG GE Compu-scene Link DIG SGI Onyx Improved rapidly “Big Iron” McDonnell Douglas VITAL Evans and Sutherland ESIG GE Compu-scene Link DIG SGI Onyx Improved rapidly, but costly… Rise of the PC McDonnell Douglas (VITAL) Evans and Sutherland (ESIG) GE (Compu-scene) Link (DIG) Improved rapidly, but VERY costly… 1999, NVIDIA GeForce 256 Rise (and fall?) of the PC Moving to the PC-based GPU brought huge cost savings Rise (and fall?) of the PC Moving to the PC-based GPU brought huge cost savings But meant giving up a lot of capability and control Real-time determinism Anti-aliasing Gen-lock Unified memory Computational power Flight Simulation vs. Gaming Instructor-controlled conditions (time, weather, etc.) Subjective tuning Fidelity No aliasing No Z-fighting No LOD popping Performance Never drop frames 20+ channels Data… Flight Simulation vs. Gaming Large gaming areas – “database” (E.g. continental, even planetary) Legacy database standards (Government-defined) Typical gaming approaches do not scale to massive datasets... (nor do traditional simulation approaches) Throw Hardware At It! Moore’s Law continues… …Via parallelism, not speed (for nearly 10 years now) First with CPU, now with GPU Not So Fast Free lunch is over… Software must be architected for parallelism: Amdahl’s Law 𝑆= 1 𝑃 (1 − 𝑃) + 𝑛 S = speed up P = % of parallelization n = number of threads The Long Pole Prepare Source Data Compile Database Re-work Intermediate Database (LODs) Load/Render Database Compute ALL the LODs! Pre-compute LODs for all possible paths into “polygon soup” Very little of the result is typically used Compute ALL the LODs! Pre-compute LODs for all possible paths into “polygon soup” Very little of the result is typically used Uses tremendous computing resources Uses tremendous amount of storage space Stuck in Time Static databases do not adapt to new data or GPU technology Stuck in Time Static databases do not adapt to new data or GPU technology Stuck in Time Static databases do not adapt to new data or GPU technology Underutilizes GPU Leveraging the GPU Large/fast memory buffers Programmability… Vertex, fragment, geometry Compute Run-time Scene Generation Prepare Source Data Source Data User-Defined Processing Rules/ Shaders Load/Render Source Data on GPU Deferred Commitment Case Study: Hardware Consolidation One 4K projector from one GPU Multiple GPUs per system Case Study: Lights ~50 tunable parameters 100’s of thousands of lights possible at 60Hz Case Study: Grass 10’s of thousands of grass blades Matches imagery color Local wind model (helo rotor wash) Summary Visualization in flight simulation continues to evolve rapidly. GPU advancements have fundamentally changed how we do things and what is possible: Improved fidelity Improved performance Deferred commitment, instant feedback Questions? Tim Woodard timw@dvcsim.com @TimAtDVC #GTC14 Thank you!