Performance analysis and workload characterization of the 3DMark05 benchmark on modern parallel computer platforms

• Authors: Fadi N. Sibai
• Venue: ACM SIGARCH Computer Architecture News - Special issue on the 2006 reconfigurable and adaptive architecture workshop
• Year: 2007

With ever increasing CPU and graphics card speeds, and improved sophistication, stunning visual effects, and growing scene detail and real life-like content of 3D games, 3DMark® emerged as the leading PC benchmark for 3D gaming performance with several millions of worldwide downloads. Its tests are at the cutting edge of consumer graphics and push the limit of 3D rendering with spectacular scenes, and state of the art lighting techniques. The benchmark scores help quickly differentiate the platforms with state of the art graphic cards and processors from those with older components. In this paper, we analyze the scaling of the 3DMark®05 benchmark with CPU frequency, number of CPUs, number of GPUs, and number of threads supported by the hardware. We also characterize the benchmark's workload. These results reveal that the benchmark scales well indicating that 3D games if implemented with multiple Physics and Artificial Intelligence or other relevant content threads should show good scaling too on multi-CPU and multi-GPU platforms. The characterization results reveal the close dependence of 3D graphics applications on the memory subsystem's performance as 1 out of 2 instructions is a load or store instruction. The results also revealed that there is a direct correlation with the Game Tests' performance and the number of cache memory read misses per instruction retired, the number of stores retired per instruction retired, the number of polygons per Draw*Primitive; and the number of setvertexshader calls per frame. All these events relate to the memory subsystem performance generally linking the 3D graphics applications' performance and the 3DMark® overall score to the platform's memory performance. Salient microarchitectural performance events of the CPU tests were also memory-related.