Gaming benchmarks may be fun, but the GTX Titan-Z is obviously primarily intended for (semi-)professional GPGPU tasks, as a "cheap" alternative to Tesla boards. We ran a number of GPGPU benchmarks on the new Titan-Z, just like we have previously done on the Titan Black, the original Titan, the GTX 780 Ti, and AMD's Radeon R9 290X.
BasemarkCL is an OpenCL benchmark which provides a measure of the GPGPU performance of a system for professional applications. The Physics tests include, among others, algorithms for 3D soft body simulations, fluid simulations, and wave simulations. The Fractal tests include algorithms for the calculation of different fractals. The Image tests use the GPU for algorithms such as colour correction, noise reduction, and sharpening. The Video tests contain comparable algorithms for video.
We can be very brief about the BasemarkCL tests: the Titan-Z has a poor showing, both as a single card and in quad-SLI.
Luxmark is a professional 3D rendering benchmark that is based on OpenCL.
The Titan-Z does impress here, and how! Both the R9 295X and the Titan are surpassed by a large margin, and for the first time we see an application where this card is actually worth the premium. Performance also scales extremely well when using two cards in a quad-SLI setup.
SiSoft Sandra 2014 contains different GPGPU benchmarks which are based on algorithms that are actually being applied in various industries. Firstly, Sandra 2014 contains a trio of benchmarks which run financial analyses according to the commonly used Black-Scholes, Monte Carlo and Binominal algorithms. Note that all of these benchmarks are based on single-precision (32-bit) floating point numbers; double-precision benchmarks (FP64) will be shown on the next page.
Unfortunately, results are once again disappointing in the Black-Scholes benchmark, with the Titan-Z being unable to keep up with the much more affordable R9 295X2. Two of these cards do have excellent performance, but compared to a single Titan Black (at less than one fifth of the cost), the peformance gain isn't remarkable.
For the Monte Carlo test, the Titan-Z demonstrates excellent scaling with respect to a Titan, and the quad-SLI Zs also perform admirably. However, the R9 295X2 stars in this benchmark, especially when considering its much lower price.
The same holds true for the Biominal benchmark, except that the R9 295X2 performs way better here.
Following this, Sandra 2014 tests a trio of algorithms which are often used in scientific analyses: matrix multiply, fast fourrier transformations (FFTs) and N-body simulations. These tests, too, are based on either integer or FP32 numbers.
We see a mixed picture, with the Titan-Z only truly flexing its muscles in the Fast Fourrier Transformations test. The N-Body Simulation results are downright disappointing, when taking into consideration the scores of the Titan and GTX models.
We zien een wisselend beeld, maar alleen in de Fast Fourier Transformaties laat de Titan-Z echt spierballen zien. De N-Body Simulation resultaten zijn ronduit teleurstellend, gezien de scores van de Titan- en GTX-modellen.
Lastly, Sandra 2014 contains a pair of cryptographic benchmarks which can leverage the GPU: the execution of AES256 encryption and the calculation of SHA2-256 hashes. The latter is, to a certain extent, comparable to the kind of work that occurs when mining for Bitcoins. We can see that AMD clearly performs better than Nvidia in this regard.
AMD still appears to be the best choice for bitcoin mining with regards to price/performance (assuming you want to use a video card, but that's another discussion entirely).