Your Tesla V100-PCIE-32GB GPU (compute capability 7.0) is not supported on CUDA 13+.

[chuksyee]

I am running on servers with these specs:

• OS: Ubuntu-24..04
• CPU: Xeon Gold 6248
• GPU: Nvidia Tesla V100-PCIE-32GB GPU .
• NVIDIA driver: Driver Version: 580.126.09 CUDA Version: 13.0

I installed Julia-1.12.5 and CUDA.jl 5.11.
For a simple test of CUDA, I tried to create a CuArray with following steps.

using CUDA
A_cpu = rand(Float32, 1024, 1024)
A_gpu = CuArray(A_cpu)

I get the following error.

julia> A_gpu = CuArray(A_cpu)
┌ Error: Your Tesla V100-PCIE-32GB GPU (compute capability 7.0) is not supported on CUDA 13+.
│ Please use a device with at least capability 7.5, or downgrade your NVIDIA driver to below v580.
└ @ CUDA /state/partition1/llgrid/pkg/julia/julia-1.12.5/local/share/julia/packages/CUDA/Il00B/lib/cudadrv/state.jl:224
ERROR: CUDA error: unknown error (code 999, ERROR_UNKNOWN)
Stacktrace:
  [1] throw_api_error(res::CUDA.cudaError_enum)
    @ CUDA /state/partition1/llgrid/pkg/julia/julia-1.12.5/local/share/julia/packages/CUDA/Il00B/lib/cudadrv/libcuda.jl:30
  [2] check(f::CUDA.var"#cuDevicePrimaryCtxRetain##0#cuDevicePrimaryCtxRetain##1"{Base.RefValue{Ptr{CUDA.CUctx_st}}, CuDevice})
    @ CUDA /state/partition1/llgrid/pkg/julia/julia-1.12.5/local/share/julia/packages/CUDA/Il00B/lib/cudadrv/libcuda.jl:37
  [3] cuDevicePrimaryCtxRetain(pctx::Base.RefValue{Ptr{CUDA.CUctx_st}}, dev::CuDevice)
    @ CUDA /state/partition1/llgrid/pkg/julia/julia-1.12.5/local/share/julia/packages/GPUToolbox/x8hVh/src/ccalls.jl:33
  [4] CuContext
    @ /state/partition1/llgrid/pkg/julia/julia-1.12.5/local/share/julia/packages/CUDA/Il00B/lib/cudadrv/context.jl:197 [inlined]
  [5] context(dev::CuDevice)
    @ CUDA /state/partition1/llgrid/pkg/julia/julia-1.12.5/local/share/julia/packages/CUDA/Il00B/lib/cudadrv/state.jl:238
  [6] CUDA.TaskLocalState(dev::CuDevice)
    @ CUDA /state/partition1/llgrid/pkg/julia/julia-1.12.5/local/share/julia/packages/CUDA/Il00B/lib/cudadrv/state.jl:50
  [7] TaskLocalState
    @ /state/partition1/llgrid/pkg/julia/julia-1.12.5/local/share/julia/packages/CUDA/Il00B/lib/cudadrv/state.jl:50 [inlined]
  [8] task_local_state!()
    @ CUDA /state/partition1/llgrid/pkg/julia/julia-1.12.5/local/share/julia/packages/CUDA/Il00B/lib/cudadrv/state.jl:79
  [9] device
    @ /state/partition1/llgrid/pkg/julia/julia-1.12.5/local/share/julia/packages/CUDA/Il00B/lib/cudadrv/state.jl:189 [inlined]
 [10] CuArray{Float32, 2, CUDA.DeviceMemory}(::UndefInitializer, dims::Tuple{Int64, Int64})
    @ CUDA /state/partition1/llgrid/pkg/julia/julia-1.12.5/local/share/julia/packages/CUDA/Il00B/src/array.jl:91
 [11] CuArray
    @ /state/partition1/llgrid/pkg/julia/julia-1.12.5/local/share/julia/packages/CUDA/Il00B/src/array.jl:437 [inlined]
 [12] CuArray
    @ /state/partition1/llgrid/pkg/julia/julia-1.12.5/local/share/julia/packages/CUDA/Il00B/src/array.jl:442 [inlined]
 [13] CuArray(A::Matrix{Float32})
    @ CUDA /state/partition1/llgrid/pkg/julia/julia-1.12.5/local/share/julia/packages/CUDA/Il00B/src/array.jl:451
 [14] top-level scope
    @ REPL[8]:1

Based on the Nvidia webpage - https://www.nvidia.com/en-us/drivers/details/260504/
under the Supported Products tab, the Tesla V100 is supported by Nvidia driver 580.126.09 and CUDA 13. So, why does Julia CUDA not work?

[JamesWrigley]

Possibly related: #3023

[vchuravy]

Got this via e-mail from a collaborator

Wondering why did CUDA.jl deprecated software support for the Volta architecture one version before NVIDIA themselves did.  NVIDIA ended support for Volta in 580, Julia ended support in 575.  580 is a Long Term Support (LTS) version that will have security patches for several years.  This puts us in an awkward position because we would be losing CUDA 13 support as well as 2 years of future security/bug fixes if we were to roll back to 575 or 570.

This is the message that we get from Julia:

+ Error: Your Tesla V100-PCIE-32GB GPU (compute capability 7.0) is not supported on CUDA 13+.
| Please use a device with at least capability 7.5, or downgrade your NVIDIA driver to below v580

Which is incorrect.  NVIDIA ended support in driver 580 (CUDA 13), not 575 (CUDA 12.9).

So yeah we likely have a simple mistake in our compatibility dictonary

[ali-ramadhan]

Tried to fix the compatibility dictionary in #3132 but support for libraries like cuFFT might be truly removed/dropped. Will investigate some more.

[maleadt]

The error message was not correct; Maxwell/Volta is still supported, but requires manual selection of an older toolkit for now. See #3023 (comment)
Error message improved in #3134

[maleadt]

I just updated the CUDA Runtime JLL to select a toolkit that's compatible with the given devices. Please test #3134.