Do you have FLOPS values?
Accurate FLOPS calculations require hardware counters that aren't available on the GPU. There are well-known FLOPS calculations for many numeric linear algebra routines, however, these values aren't completely valid when retooled for GPU acceleration.
Where can I find benchmarks for individual routines?
We have organized more specific benchmarks by their family of routines:
» Linear Equation Solvers
» Orthogonal Factorizations
» Least Squares Solvers
» Symmetric Eigenproblems
» General Eigenproblems
» Singular Value Decomposition
This graph shows the relative speed up of many CULA functions when compared to the Intel's MKL 10.2 implementation of LAPACK.
Please note: Complex, Double, and Double Complex performance will be posted soon!
All benchmarks were obtained using:
CPU: Quad-core Intel Core i7 @ 2.8 GHZ CPU
GPU: NVIDIA Tesla C1060
OS: Windows 7 (64-bit)
CULA speed calculations include the data transfer time to and from the GPU.
MKL speed calculations were obtained with all cores and hyper-threading active.