CUDA device support

Specific Domains GPU
1

I have an old Nvidia 610M GPU on my laptop. It is not a powerful GPU but I wanted to play around and learn a bit of GPU computing. Although the GPU supports CUDA (I have version 9.1.112), I seem to encounter some problems.

When testing the CUDA package, the following error is thrown:

(@v1.4) pkg> test CUDA
    Testing CUDA
...
ERROR: LoadError: Could not find any suitable device for this configuration

And when following the tutorial for CUDA.jl:

julia> N = 2^20
1048576

julia> x_d = CUDA.fill(1.0f0, N)  # a vector stored on the GPU filled with 1.0 (Float32)
1048576-element CuArray{Float32,1,Nothing}:
 1.0
 1.0
 1.0
 1.0
 1.0
 1.0
 ⋮
 1.0
 1.0
 1.0
 1.0
 1.0

julia> y_d = CUDA.fill(2.0f0, N)
1048576-element CuArray{Float32,1,Nothing}:
 2.0
 2.0
 2.0
 2.0
 2.0
 2.0
 ⋮
 2.0
 2.0
 2.0
 2.0
 2.0

julia> y_d .+= x_d
ERROR: Device capability v2.1.0 not supported by available toolchain

Is that because my GPU is not supported, or my CUDA version? Is there a list of supported GPUs?

2

https://developer.nvidia.com/cuda-gpus CUDA 2.1 is wayyy too old

Edit: this is capability not compatibility, for 610m, it supports till CUDA 8

3

Capture
Capture558×516 16.2 KB

4

you can install DLL all you want. Doesn’t mean your device support it

5

Interesting, I didn’t install v9 myself. Why would Nvidia choose then these drivers? Might installing v8 help? Or is it also unsupported?

6

It seems CUDA.jl supports only CUDA v9 and above.

julia> x_d = CUDA.fill(1.0f0, N)
┌ Warning: CUDA.jl only supports NVIDIA drivers for CUDA 9.0 or higher (yours is for CUDA 8.0.0)
1 Like
7

If you upgrade your driver to have CUDA capability 9.0 or higher, CUDA.jl should automatically download a matching version of the toolkit. Although your screenshot above already mentions CUDA 9.1; could you try running with JULIA_DEBUG=CUDA? What does CUDA.version() show?

8

I installed CUDA 8.0 to test it, now I tried to install back a higher version but NVIDIA assigned and installed again CUDA 8.0. I don’t know why or how I had CUDA 9.1 installed since it shouldn’t support Fermi GPUs.

julia> CUDA.version()
┌ Warning: CUDA.jl only supports NVIDIA drivers for CUDA 9.0 or higher (yours is for CUDA 8.0.0)
└ @ CUDA C:\Users\Martin Cornejo\.julia\packages\CUDA\5t6R9\src\initialization.jl:109
ERROR: Could not find a suitable CUDA installation

I should have tried this before changing the drivers…