BIDData/BIDMach

Build against CUDA 8.0 / Support GPU Only Builds?

Opened this issue · 5 comments

dakami commented

./getdevlibs.sh doesn't find any 8.0 libraries, and the icc dependency makes it impossible to rebuild locally. Might be nice to have a GPU only build path, but would definitely be nice if I could build against CUDA 8 / these Titans.

svanschalkwyk commented

Any development on this?

DanielTakeshi commented

@svanschalkwyk @dakami I've successfully used CUDA 8.0 on my machine (with Titan GPU with Pascal). Did you use the maven update? (It's on the Wiki page right now on this GitHub).

svanschalkwyk commented

It is only seeing one of the two CUDA devices. I have export CUDA_VISIBLE_DEVICES=1,2

svanschalkwyk commented

I downloaded the pre-built binaries and they also show only one device.
"1 CUDA device found, CUDA version 7.0"

Wed Feb 8 12:43:07 2017
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 375.26 Driver Version: 375.26 |
|-------------------------------+----------------------+----------------------+
| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
|===============================+======================+======================|
| 0 GeForce GTX 960 Off | 0000:04:00.0 On | N/A |
| 0% 25C P8 7W / 160W | 87MiB / 4032MiB | 0% Default |
+-------------------------------+----------------------+----------------------+
| 1 GeForce GTX 760 Off | 0000:83:00.0 N/A | N/A |
| 42% 32C P8 N/A / N/A | 32MiB / 4036MiB | N/A Default |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Processes: GPU Memory |
| GPU PID Type Process name Usage |
|=============================================================================|
| 0 2340 G /usr/lib/xorg/Xorg 66MiB |
| 0 2861 G cinnamon 19MiB |
| 1 Not Supported |
+-----------------------------------------------------------------------------+

svanschalkwyk commented

[SOLVED]
CUDA_VISIBLE_DEVICES have to follow the numbering that nvidia-smi shows, i.e. 0 for GPU 0, 1 for GPU 1.
Suggestions (somewhere else) that the GPU array is 1-based are incorrect.