Nvidia GeForce GTX 780 Ti Review: GK110, Fully Unlocked 88
An anonymous reader writes "Nvidia lifted the veil on its latest high-end graphics board, the GeForce GTX 780 Ti. With a total of 2,880 CUDA cores and 240 texture units, the GK110 GPU inside the GTX 780 Ti is fully unlocked. This means that the new card has an additional SMX block, 192 more shader cores, and 16 additional texture units than the $1,000 GTX Titan launched back in February! Offered at just $700, the GTX 780 Ti promises to improve gaming performance over the Titan, yet the card has been artificially limited in GPGPU performance — no doubt in order to make sure the pricier card remains relevant to those unable or unwilling to spring for a Quadro. The benchmark results simply illustrate the GTX 780 Ti's on-paper specs. The card was able to beat AMD's just-released flagship, the Radeon R9 290x by single-digit percentages, up to double-digits topping 30% — depending on the variability of AMD's press and retail samples."
Re:heh (Score:5, Insightful)
No, you were right the first time. Toms Hardware is the only site claiming these benchmark victories for the nVidia card. I'm not saying they allow their advertising department to influence their reporting and rankings, but it's a bit fishy that they're such an outlier regarding the flagship video cards of the two manufacturers.
It's also worth noting that comparing these cards without taking AMD's Mantle technology into account is to say the least, incomplete.
So stop using CUDA? (Score:4, Insightful)
Re:So stop using CUDA? (Score:5, Insightful)
Ironically, in the industry I work in (computer vision, embedded signal processing, etc) - we've slowly been moving AWAY from OpenCL - it's dead/stagnant, still hasn't caught up to where CUDA was 3 years ago, and to put this in general terms, there's fragmentation among both support of STANDARD (as in, the specification) features of OpenCL and worse that prevent you using simple things like images or barriers properly.
On top of that, radically different implementations obviously have radically different optimization processes - for one particular kernel, we were looking at hitting 13 different optimizations of the same kernel/function, to target different devices.
Now days, we use CUDA - compiled to PTX (for nVidia, x86_64, and ARM NEON) and CAL for legacy AMD, HSAIL for GCN AMD, and GLSL for Intel.
Say what you will about CUDA/PTX vendor lock in, PTX is far more device agnostic than OpenCL is, just as portable (thanks to great open source efforts like LLVM (and the vendor support nVidia/AMD/Intel/etc provide) and GPUOcelot), and far more mature than HSAIL is.
That's by basis for NOT using OpenCL, where's yours?