Benchmark Results: Sandra 2012 And LuxMark 2.0
For all of their alacrity in gaming, the Kepler-based GeForce products demonstrate an evolutionary step forward in FP32 throughput and a more jarring slow-down in double-precision floating-point math. Meanwhile, AMD’s GCN architecture propels Tahiti forward in compute workloads.
Sandra 2012 helps quantify AMD’s gains and Nvidia’s willing compromise. The only cards GeForce GTX 670 beats in floating-point math are older Fermi-based boards. Meanwhile, the new 670 takes last place in FP64.
Leaning on this diagnostic tool again shows us bandwidth between the GPU and its memory, along with throughput over the PCI Express bus. Extremely low data transfer numbers across the board represents poor performance from the X79 platform. Meanwhile, internal bandwidth reflects on the 384-bit aggregate memory bus AMD uses on its Radeon HD 7900-series cards, along with the 256-bit interface Nvidia’s employs on its GeForce GTX 600s. GeForce GTX 690 seems to break this test, returning a bad GP call.
Applying those theoretical numbers from Sandra to a real-world-based test like LuxMark paints a pretty telling picture. AMD’s cards absolutely dominate. The slowest, Radeon HD 6990, even manages to beat the fastest Nvidia board—ironically, the old GeForce GTX 590.
Truly, we’re interested to see how Nvidia plans to address compute on the desktop after making such a big deal about it last generation. Might the company have plans to sell you a separate compute-oriented upgrade?