GPUPI 2.3 is Launched, Includes Optimizations for Cuda 8.0

GPUPI 2.3 is Launched, Includes Optimizations for Cuda 8.0

The latest 2.3 version of GPUPI is now available from where Mat has managed to include several optimizations that will be good news for anyone benching an Nvidia card including the latest Pascal-based 1XXX series. Older cards should also see an improvement, for example a GTX 750 Ti has seen performance of up to 10 seconds faster than with version 2.2. As well as these performance improvements there are also several other changes, outlined below in the following notes:

    Validation of partial results during the run: In GPUPI 2.2 and before only the final result was validated. The new version checks after every 4th loop, if the score is still alright. If an error occurs, the benchmarks stops. That should help especially with old or slow hardware to know as fast as possible the system is not stable/able to produce a good result.

    Validation of result files: You can check your sandbagging result files for validity now, so no tears any more when you upload your backup for a competition just to see that the file is corrupt.

    DPI Awareness ... or let's call it 4K ready: People tend to use 150% on smaller 4K monitors, which leads to wrong window sizes or even worse blurred fonts. GPUPI is now aware of the DPI settings the desktop uses and scales the fonts and windows for you.

    Improvements for the output: Added theoretical occupancy for CUDA kernels, memory allocation is now printed as MB instead of Bytes.

    The Legacy version: is now compiled with Visual Studio 2013 Update 5, so it has the same dependencies as the normal version of GPUPI. An important note here: I've seen people using this version for Windows 10 and newer graphics cards. Don't do this, this version is only for Windows XP and/or GeForce 200 series cards and it's definitely the slower choice and the Legacy version can not handle 4K screenshots due to it's older Windows libraries.

    Download the latest 2.3 version of GPUPI here at

Please log in or register to comment.