You're right -- this paper isn't particularly helpful to anyone who isn't already steeped in GPU technology. Here's most of what you need to know: "SMX: 192 single-precision CUDA cores, 64 double-precision units, 32 special function units (SFU), and 32 load/store units (LD/ST)." If you're doing 32-bit single precision, I think you basically have an SIMD array of 192 processors. "CUDA" is nVidia's C-based language for programming these things. Since Mandelbrot does a reasonable amount of processing per pixel, it shouldn't be memory-bound. Even GPU memory is getting pretty cheap, so you can have a very large image store -- hence the ECC on the image memory. At 03:23 PM 8/18/2012, Fred lunnon wrote:
For all the sense I could make of it, this document might just as well have been an extract from a storyboard for Star Wars --- and I used to teach computer architecture a few years ago!
WFL
On 8/18/12, Henry Baker <hbaker1@pipeline.com> wrote:
Can't modern GPU's (nVidia Kepler, etc.) compute Mandelbrot in "real time" -- i.e., the time it takes to display that portion of the screen ? If not, aren't modern GPU's getting pretty close to this capability ?
www.nvidia.com/content/PDF/kepler/NVIDIA-Kepler-GK110-Architecture-Whitepaper.pdf