70 likes | 235 Views
GPGPU and CUDA. Center for Visual Information Technology, IIIT Hyderabad. Singular Value Decomposition on GPU. Sheetal Lahabar. 2 Step SVD bidiagonalization on GPU and hybrid CPU/GPU implementation for Diagonalization Error due to lower precision < 0.001% Upto 8x faster than Intel MKL
E N D
GPGPU and CUDA Center for Visual Information Technology, IIIT Hyderabad
Singular Value Decomposition on GPU Sheetal Lahabar • 2 Step SVD bidiagonalization on GPU and hybrid CPU/GPU implementation for Diagonalization • Error due to lower precision < 0.001% • Upto 8x faster than Intel MKL • Upto 59x faster than MATLAB
Artificial Neural Networks on GPU Sheetal Lahabar • ANN training and classification, batch learning formulated in CUBLAS • Average classification time for 1K test pattern is 0.759 sec • Upto 210 times faster than MATLAB • Upto 267 times faster than FANN