uBench is a suite of micro-benchmarks for GPUs (Graphic Processing Units), developed using Nvidia CUDA programming model.

It is designed to test the impact on performance of kernel launching parameters (mainly thread-block sizes/geometries, and L1 cache configuration). Each micro-benchmark test different application scenarios, testing different global memory access patterns, ratios of arithmetic operations vs. global memory accesses, etc. A description of the benchmark suite can be found in the paper:

uBench: Exposing the Impact of CUDA Block Geometry in Terms of Performance.
Y. Torres, A. Gonzalez-Escribano, D.R. Llanos.
The Journal of Supercomputing, 65 (3), pag. 1150-1163.
Springer, 2013.

Please, reference this paper if you need to cite uBench.

uBench v1.0

The uBench v1.0 suite, including the experimentation results presented in the previous paper.

Dowload: uBench_v1.0,  .tgz, 9,8 KB

uBench v1.1_tc

The uBench v1.1_tc suite is a modified version of uBench v1.0 specialized for 1D arrays.

Contribution of Christopher W Quackenbush (New York University).

Donwload: uBench_v1.1, .zip, 6,1 KB