^ What's the point?
^ Quick start
^ Notes

Download :: Top
GFX Memory Speed Benchmark

A free GFX RAM and system RAM speed benchmark application. Measures GFX card RAM and system RAM read and write access speed by the CPU, basically the graphic card's RAM accessed by CPU through the PCIe bus and system RAM's access speed by the CPU.
Displays GFX RAM and system RAM average and peak read and write speeds.
OpenCL RAM benchmark measures combined read-then-write performance, also usable to benchmark OpenCL supporting compute cards/devices.
Note that the results should be compared only between the same build: Win32 with Win32 or Win64 with Win64 and with the same version of GFX Memory Speed Benchmark.
Results are relative to the tests performed by the app., they are not absolute values, compare the results to other benchmarks performed by the app. (eg. on other system, or for example when upgrading your computer with a new GPU, CPU, RAM, or with a new computer).
GFX Memory Speed Benchmark is free. If you like this software please consider registering one of my shareware applications.

The Delphi source code is available for €25, although no re-selling and redistribution allowed, only making custom builds for personal or company use or for educational purposes.

Quick start

Start the app. and select the platform of the OpenCL device (eg. the graphics card's, for example 'NVIDIA'), select the actual graphics card ('Device') of that vendor to the right. Important step is to select the "Compute core count" of the device, there are some presets to choose from. If your card is not listed, try to choose a preset that is the closest match. If you try multiple presets the highest result value is the most true. Also it's possible to just enter a value, Google for your device's real (eg. CUDA cores in case of Nvidia) value.
Click on "Run benchmark", the results should be displayed after the benchmarking.


Always specify at least 2-3 passes. A little delay is before "turbo boost" kicks in, so the first pass usually reports a lower result than the actual real speed. The second pass should run with full processing speed.