GTC 2020: HPL-AI: Benchmarking Half-Precision Hardware with Modern Numerical Linear Algebra
After clicking “Watch Now” you will be prompted to login or join.
Click “Watch Now” to login or join the NVIDIA Developer Program.
HPL-AI: Benchmarking Half-Precision Hardware with Modern Numerical Linear Algebra
Piotr Luszczek, University of Tennessee
We'll present the mixed-precision iterative and direct methods used by the HPL-AI benchmark. These new approaches are instrumental in kernel-based performance evaluation of modern hardware accelerators that offer fast implementation of limited-precision floating-point units. The scope of the benchmark spans a number of important computational patterns that scientific codes, both in the past and on modern GPUs, often rely upon. We had to resolve a number of numerical issues to achieve a robust implementation, and we'll present those results for relevant background on the design process involved in benchmarking and how it can be leveraged for the benefit of the GTC audience.