News of NVidia’ GTX580 is all over the internet today, but one of the better reviews comes from the Jon Peddie Report who takes it through a series of popular benchmarks. So what makes the card so great? Well, compared to the 480:
Now all 512 cores and all 16 polymorph engines are running, and at a higher clock with the same or less wattage than the crippled 480. Super good tessellation, great ROPS, and 192 GBytes/sec high-speed memory transfers.
They took the card and pitted it against the GTX480, along with the AMD Radeon HD5870 and 6870 in a collection of games, and ranked them (You should like this Jeff) by Framerate.
But that’s not the full story. The GTX480 easily beats all four cards in raw framerate, but only barely squeaks past the 6870 in Power Per Watt, and loses in Performance Per Dollar. In the end, the card winds up being the flat-out-winner in raw horsepower, but possibly not the best use of limited funds. Of course, all of these tests were performed on games, and not engineering applications. I’ld personally like to see some SpecViewPerf results, and I’m very curious to see how some CUDA codes would run with the additional cores enabled. In our Quadro5000 review we showed that Raw Core-count is not a good indicator of performance, as the Quadro’s could easily beat the GeForces, even though had a lower CUDA-core count.
Definitely read the review for all the gory details and graphs.
via Nvidia GTX580 Review – Mount Tiburon Testing Labs Reviews.
I never like the “value” numbers. Sure the card costs more, but compared to what? Is it more than another card? Is either card useful without a computer? Is that computer useful without software? Is that software useful without an operator? Add up the cost of the card, the computer, the software licenses, and my salary and benefits and the overhead of our operations and see how much I care about what ends up being .01% difference in cost, but a 3% improvement in productivity.
I must say though, having only 1.5GB in the GTX580 is a letdown. Why do mobile Fermi parts have more memory than desktop parts?