Is it true that the double precision performance for GTX690 is only 1/24 of it’s single precision performance? What about the Tesla K20?
Yes, GTX690 DP is 1/24 of it’s SP counterpart. Tesla K20 and K20X have DP performance of 1/3 it’s SP counterpart. GTX Titan has 1/24th of it’s SP by default, but can be switched via the NVidia configuration to use the DP units to net the same 1/3 DP performance compared to its SP counterpart.
Why cannot GTX 690 switch?
Because it doesn’t have any more DP units that could be switched on.
How is this done? I this a flag in the compiler or something in the driver?
pasoleatis, the DP mode can be changed with nvidia-smi or with the NVML library. Look “GPU Operation Mode (GOM)” in the docs. There’s one option for enabling/disabling fullspeed DP, and one option for enabling/disabling graphics support.
NVIDIA’s drivers restrict GOM on K20c (it’s reserved for K20m and K20x). But anecdotally, Titan supports GOM (online reviews show the DP mode switch in the driver screen in Windows at least).
actually nvidia-smi doesn’t support GOM switch on Titan, otherwise we wouldn’t have this thread:
[url]Turning on DP for Nvidia Titan on Headless Server? - Linux - NVIDIA Developer Forums
The place it can be changed is in NVIDIA Control Panel (Windows) and nvidia-settings (Linux)
Thanks for your replies. I have a pair of 660 TI and it appears there is still space on the mother board and in the case (1500 W power source) for a Titan gtx. Most of my problems require double precision and I just wanted to check before I spend another 1000 euros.