Thread: GPU ray tracing benchmark on Radeon HD 7970 and GeForce GTX

Original articles from 30 January 2012: The Luxmark was chosen in his first version of AMD as a showpiece for the launch of the Radeon HD 7970 and should be focused on the particular suitability of the graphics core architecture for next-GPU compute the proper light. In actuality, the Radeon in the standard run, the predecessors like the GeForce competition by a factor of two or more dependent. 1.0 Luxmark used by default still a very modest scene is Luxmark 2.0 with three benchmark versions on the download servers from which the sala scene is of medium complexity is the default. A beta version of the test we have in PCGH 03/2012, which from 1 February in the trade, is already used to compare the GPU compute power of current graphics cards.


Ray tracing with the graphics card is an ideal case of GPU Compute: Massively parallel computations are hundreds to thousands of arithmetic units meet modern GPUs, but demand in the same breath, the interconnection of the various execution units to their utmost. Because the GPU is really fast only if the shader ALUs are both lined with timely data and the caching effects of the cache server can be exploited. The more complex the scene, and the more so called Light bounces, reflections of light rays, therefore, be included in the calculation, the higher the claims not only to the shader processing plants but also and especially to caches and memory controllers.




We looked at our current VGA test system with an overclocked Core i7 CPU, the performance of today's GPUs, the performance and high-end range and comparing both the Intel CPU with the latest AMD OpenCL driver and two older high-end cards in the form of the Radeon HD 4890 and GeForce GTX 285 also assayed. For a comparison of the OpenCL driver, please refer to the PCGH 03/2012, which from 1 February in the trade - where we have also assayed and the Intel OpenCL driver.


As already mentioned, the 2.0 Luxmark has three benchmark scenes: Sala (standard, medium complexity), room (high complexity, only 64-bit systems) (if low complexity, even in Luxmark 1.0) and Luxball HDR. PC Games Hardware has tested all three benchmarks available in the 64-bit version in order to document the effects of increasing scene complexity on the GPU suitability.


An interesting observation we made with the Nvidia driver. Since we used the Luxmark before about three quarters of a year in issue 09/2011, we had a data set that made us sit up in comparison with the current values. With the older drivers 275.33er reached and reach the Geforce cards much better values ​​than with current drivers - whether in the meantime was switching from OpenCL 1.0 to 1.1 in connection with it stands or whether it is a simple bug in the driver concerned, could Nvidia us to explain, despite repeated request not to date.