According to Hexus, the desktop version of the GF100 GPU, which powers the GeForce GTX 480 / GTX 470, has a limited double precision speed. Indeed, the double precision speed is limited to 1/8 of single precision speed while on the HPC version of the Fermi architecture (HPC = high performance computing), the double precision speed is limited to 1/2 of single precision speed.
On Radeon HD 5000 series, the double precision speed is limited to 1/5 of single precision speed (source):
– single precision: 2720 GFlops
– double precision: 544 GFlops
For the sake of memory, GF100 / Fermi architecture implements the new IEEE 754-2008 floating-point standard, that offers full 32-bit and 64-bit precision for all instructions.
Actually I think that rv870 and rv770 provide 1/5th of the single precision, not 1/2nd!
5870 – 2.7 TeraFLOPS single-precision and 544 GigaFLOPS double-precision performance
Tesla (672 Gflops double) > 5870 (554Gflops double) > GTX 480 (168 Gflops double)
Thanks guys, I update the post right away.
The Fermi Tesla C2070/C2050 are rated at 630/520 Gflops DP peak.
Seems like adding 64bit DP to OpenGL 4 was a mistake! First we find out that low end 5000 series will not run DP natively, now Fermi has been crippled.
At least for this GPU generation DP is going to be way to slow for graphics and games.
Pingback: 2-GTX460 in SLI or 1 GTX480? - Page 3 - Overclock.net - Overclocking.net