CUDA 6.0 and Titan-Z


I see in debian’s library it has at most CUDA 6.0 which was released march 2014

The GPU i bought is a Titan-Z also released march 2014. Does CUDA 6.0 support the Titan-Z correctly?
My intentions is to run the linux box in textmode. So i only need to be able to develop correctly from textmode console (remote SSH). Yet of course debugging and profiling my code is important feature.

/home/diep# apt-cache search cuda
boinc-nvidia-cuda - metapackage for CUDA-savvy BOINC client and manager
libeztrace0-contrib - Automatic execution trace generation for HPC - development files
libhwloc-contrib-plugins - Hierarchical view of the machine - contrib plugins
python-pycuda - Python module to access Nvidia‘s CUDA parallel computation API
python-pycuda-dbg - Python module to access Nvidia‘s CUDA API (debug extensions)
python-pycuda-doc - module to access Nvidia‘s CUDA computation API (documentation)
python3-pycuda - Python 3 module to access Nvidia‘s CUDA parallel computation API
python3-pycuda-dbg - Python 3 module to access Nvidia‘s CUDA API (debug extensions)
libthrust-dev - Thrust - Parallel Algorithms Library
pyrit - GPGPU-driven WPA/WPA2-PSK key cracker
pyrit-opencl - OpenCL extension module for Pyrit
python-pytools - big bag of things supplementing Python standard library
python3-pytools - big bag of things supplementing Python 3 standard library
python-pyviennacl - Python bindings for ViennaCL linear algebra library (for Python 2)
python3-pyviennacl - Python bindings for ViennaCL linear algebra library (for Python 3)
pyviennacl-doc - HTML documentation for PyViennaCL
suricata - Next Generation Intrusion Detection and Prevention Tool
vim-syntastic - Syntax checking hacks for vim
libcublas6.0 - NVIDIA cuBLAS Library
libcudart6.0 - NVIDIA CUDA Runtime Library
libcufft6.0 - NVIDIA cuFFT Library
libcufftw6.0 - NVIDIA cuFFTW Library
libcuinj32-6.0 - NVIDIA CUINJ Library (32-bit)
libcuinj64-6.0 - NVIDIA CUINJ Library (64-bit)
libcupti-dev - NVIDIA CUDA Profiler Tools Interface development files
libcupti-doc - NVIDIA CUDA Profiler Tools Interface documentation
libcupti6.0 - NVIDIA CUDA Profiler Tools Interface runtime library
libcurand6.0 - NVIDIA cuRAND Library
libcusparse6.0 - NVIDIA cuSPARSE Library
libnppc6.0 - NVIDIA Performance Primitives core runtime library
libnppi6.0 - NVIDIA Performance Primitives for image processing runtime library
libnpps6.0 - NVIDIA Performance Primitives for signal processing runtime library
libnvblas6.0 - NVBLAS runtime library
libnvvm2 - NVIDIA NVVM Library
nvidia-cuda-dev - NVIDIA CUDA development files
nvidia-cuda-doc - NVIDIA CUDA and OpenCL documentation
nvidia-cuda-gdb - NVIDIA CUDA Debugger (GDB)
nvidia-cuda-toolkit - NVIDIA CUDA development toolkit
nvidia-nsight - NVIDIA Nsight Eclipse Edition
nvidia-profiler - NVIDIA Profiler for CUDA and OpenCL
nvidia-visual-profiler - NVIDIA Visual Profiler for CUDA and OpenCL
libcuda1 - NVIDIA CUDA Driver Library
libcuda1-i386 - NVIDIA CUDA 32-bit runtime library
libnvcuvid1 - NVIDIA CUDA Video Decoder runtime library
libnvidia-compiler - NVIDIA runtime compiler library
nvidia-cuda-mps - NVIDIA CUDA Multi Process Service (MPS)
nvidia-opencl-icd - NVIDIA OpenCL installable client driver (ICD)

Kind Regards,
Vincent Diepeveen

Titan Z is kepler class hardware. CUDA 6 was the first CUDA release with explicit support for Kepler GPUs. So in general it should be possible to use the 2 together. The main thing to ensure functionality would be to use a proper GPU driver for the Titan Z. A driver newer than the one that originally shipped with CUDA 6 may be necessary (is probably necessary). I have no idea what GPU drivers may be in debian’s library. In general I don’t recommend using distro-provided libraries for these things, since you don’t know how a 3rd party packaged things. I personally would suggest using NVIDIA provided packages (or runfile installers) for CUDA and for the GPU driver.

Many thanks for your comments - A runfile that works for DEBIAN with cuda 6.5 would be really great - yet debian weirdly doesn’t get supported by Nvidia - Furthermore version 6.5 as i see a report posted that cuda 7.0 has less double precision powers by default and i’m gonna code a double precision FFT transform for it (which also will work at newer Tesla’s great just i can’t afford those).

If double precision performance by default works on a newer CUDA versions works ok and there is a Debian runfile i would definitely install that one.

I’m settling for version 6.0 and will have a look whether it works fine.

Is a driver really needed?

If i do not use it as a graphics card yet only as a CUDA device - i boot the box on the SGI card, do i really need a driver then installed?

As after i install the driver it’ll automatically boot from the Titan-Z card as device-0 and i want to use it 100% to crunch numbers, not as a display card :)

Hello if i install display driver 365.19 will that work for CUDA 6.0?

I found 347.88 as the ‘oldest’ driver available there on website nvidia.
Downloading that one now.

Hopefully is not CUDA 7.0!

yahoo found older one: 344.60 now for CUDA 6.5 - hope that one works for 6.0.37 also (all windows Vista ultimate x64)

with 344.60 installed - then 6.0.37 cuda doesn’t even ask for installing itself - just doesn’t seem to install itself.

Ok next attempt is gonna be OpenSuse 13.1 + CUDA toolkit 6.5 + Nvidia driver (says it’s for cuda 6.5) - that is for tomorrow (which is a few hours from now as it’s 5 AM now).

Hopefully winning combination!