Event Type:

Applied Mathematics and Computation Seminar

Date/Time:

Friday, June 7, 2019 - 12:00 to 13:00

Location:

STAG 160

Event Link:

Guest Speaker:

Timothy Costa

Institution:

NVIDIA CUDA Math Libraries

Abstract:

NVIDIA V100 GPUs have a peak double precision performance of 7.8 trillion floating point operations per second (TFLOPS), and a peak single precision performance of 15.7 TFLOPS. However, when leveraging the GPU's tensor cores the V100 has a peak half precision performance of 125 TFLOPS. Thus a natural question arises; can high performance computing applications leverage the immense half precision computational power of these GPUs without losing the double precision accuracy required in many scientific computing domains? The answer turns out to be yes, and in this talk we examine upcoming features in the cuSOLVER library which leverage tensor cores to accelerate the solution of dense linear systems by LU factorization while maintaining double precision accuracy.

BIO: Timothy Costa completed his Mathematics PhD at OSu in 2016. After working at Intel in MKL group, since February he is Product Manager at CUDA Math Libraries at NVIDIA.