OpenCL: GPGPU Benchmarks
Overall, OpenCL-based tests are far less friendly to Nvidia's desktop GK104-based cards than the games. In order to add more data, we're including the workstation-oriented products we just finished testing on the same machine, as well as previous-generation desktop boards. Depending on the application, its workload, and driver optimizations, the outcome can change dramatically. There are still a handful of obvious trends, though.
We list three different settings for the GeForce GTX 660 Ti, which correspond to clock rates we're told will be available at retail. Incidentally, we're currently working on a round-up of GTX 660 Ti cards, and all three speeds will appear there as well. We also re-tested AMD’s reference Radeon HD 7950 flashed with the "GHz Edition" firmware.
AMD’s consumer and workstation cards dominate the field in Bitcoin mining and SHA-256 hashing. Clearly, if this is a workload that keeps your graphics hardware hammered all day and night, the GeForce GTX 660 Ti is not for you. Even the most aggressive overclock sees the new board fall short of Nvidia's old Fermi-based GeForce GTX 580.
We purposely chose LuxMark 2.0's simpler rendering workload. Otherwise, the weaker cards from Nvidia would have given us even shorter bars, making them indiscernible. However, the GeFore GTX 570 and 580 enjoy more pronounced leads as the scene's complexity increases.
The PostFX test combines OpenCL and OpenGL in one demanding workload. Although this benchmark is taken straight from Nvidia’s SDK, it’s also an impressive example of how quickly the performance story can change in the fact of a new architecture like GCN.
Finally, Nvidia's Kepler-based cards exhibit a stronger showing. But again, it's the company's previous-gen architecture that comes out on top.