I have hardware (and associated driver) that can only work with CUDA 6.5
The installation instructions for CUDA.jl state that CUDA 10.1 is required for the current version, but that earlier versions may work with earlier versions of CUDA. I wonder how one might go about trying this out, and if anyone has succeeded?
Maybe. What I found in the source code, even on master:
elseif err.code == CUPTI_ERROR_CUDA_COMPILER_NOT_COMPATIBLE
"Profiling results might be incorrect for CUDA applications compiled with nvcc version older than 9.0 for devices with compute capability 6.0 and 6.1"
elseif err.code == CUPTI_ERROR_LEGACY_PROFILER_NOT_SUPPORTED
"an unknown internal error has occurred. Legacy CUPTI Profiling is not supported on devices with Compute Capability 7.5 or higher (Turing+)"
I had never tried CUDA (locally) and for the first time installed (it takes its time, 10+ minutes downloading 2GB+), and this is the amount of testing I did:
$ nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2017 NVIDIA Corporation
Built on Fri_Nov__3_21:07:56_CDT_2017
Cuda compilation tools, release 9.1, V9.1.85
julia> @time using CUDA
[ Info: Precompiling CUDA [052768ef-5323-5732-b1bb-66c8b64840ba]
78.977272 seconds (15.36 M allocations: 915.469 MiB, 0.52% gc time)
I would expect some stuff breaking at runtime, but I haven’t looked into what got added with CUDA 10, or specific compute capability, and maybe you do not need it?
You could try (if you get an error right there or later):
pkg> add CUDA@v1.3.3 # I didn't specify version, and I don't know what this is supposed to support.
Thanks! I’ll give it a try.