"The G in GPU is for Graphics damnit "
4 days ago
- #Physarum Simulation
- #Triton DSL
- #GPU Optimization
- The author worked with a graphics researcher who was frustrated by GPUs being used for ML workloads instead of graphics, leading to a compromise on working with NeRFs.
- Developed an interest in 'code art' during the first year, admiring works by fronkonstin, Sage Jenson, and others, and experimented with Processing and TouchDesigner.
- Joined Microsoft Research India for a Bachelor’s thesis, working on systems and GPU optimizations, learning about Triton, a DSL for CUDA programming.
- Explored Physarum simulation, a model of mold growth using agent and pheromone fields, implemented in PyTorch with performance analysis.
- Discussed the sensitivity of the Physarum model to initial conditions, drawing parallels to pseudo-random algorithms and boolean function analysis.
- Profiled the performance of the Physarum simulation, identifying bottlenecks and optimizing with Triton for faster execution.
- Demonstrated the use of Triton to write efficient CUDA kernels, improving performance by reducing kernel launches and memory usage.
- Shared visual results of the Physarum simulation, including different initializations and color palettes.