The challenge of managing parallelism and data-movement for tensor computations on GPUs
summary
Claims that dense tensor computations like matrix multiplication and convolution exhibit very regular, structured parallelism so finding a load-balanced distribution across thousands of GPU cores is essentially straightforward. The real challenges are controlling data movement through the GPU memory hierarchy and ensuring performance models actually match measured throughput, not the irregular-sparse modeling problems; the talk contrasts dense-GPU behavior with irregular algorithms and emphasizes where optimization effort must go rather than on balancing work.
AI-written summary · about
chapters (3)
related talks
50mGeometry on Steroids: Practical Experience Porting Heavy Computation to GPU
01h 21mPanel+Closing
27mConsistency and Coherence of the NVIDIA Grace-Hopper Superchip
49mGraph Based Processing of Big Images
32mChallenges in using accelerators via programming languages
10mA Programming Model for Efficient Inter-Kernel Control-Flow on Memory-Mapped Near-Data(…)
33mManaging High-Bandwidth Memory is a Parallel Scheduling Problem
47mBest Practices for Real-time Intelligent Video Analytics