RTX5070Ti is not fully supported by CUDA 12.8.0 #2714

Yongda1 · 2025-03-23T03:15:51Z

When i am running the following codes,

import Pkg
using CUDA
using Test
Pkg.add("CUDA")
CUDA.versioninfo()
N = 100
x_d = CUDA.fill(1.0f0, N)  # a vector stored on the GPU filled with 1.0 (Float32)
y_d = CUDA.fill(2.0f0, N)  # a vector stored on the GPU filled with 2.0
y_d .+= x_d
@test all(Array(y_d) .== 3.0f0)

Error

1 device:
  0: NVIDIA GeForce RTX 5070 Ti (sm_120, 13.432 GiB / 15.921 GiB available)
┌ Warning: Your NVIDIA GeForce RTX 5070 Ti GPU (compute capability 12.0) is not fully supported by CUDA 12.8.0.
│ Some functionality may be broken. Ensure you are using the latest version of CUDA.jl in combination with an up-to-date NVIDIA driver.
│ If that does not help, please file an issue to add support for the latest CUDA toolkit.
└ @ CUDA C:\Users\Administrator\.julia\packages\CUDA\sWPBr\lib\cudadrv\state.jl:236 
ERROR: 
Stacktrace:
  [1] error(s::String)
    @ Base .\error.jl:35
  [2] CuModule(data::Vector{UInt8}, options::Dict{CUDA.CUjit_option_enum, Any})
    @ CUDA C:\Users\Administrator\.julia\packages\CUDA\sWPBr\lib\cudadrv\module.jl:69
  [3] CuModule
    @ C:\Users\Administrator\.julia\packages\CUDA\sWPBr\lib\cudadrv\module.jl:49 [inlined]
  [4] link(job::GPUCompiler.CompilerJob, compiled::@NamedTuple{image::Vector{UInt8}, entry::String})
    @ CUDA C:\Users\Administrator\.julia\packages\CUDA\sWPBr\src\compiler\compilation.jl:414
  [5] actual_compilation(cache::Dict{…}, src::Core.MethodInstance, world::UInt64, cfg::GPUCompiler.CompilerConfig{…}, compiler::typeof(CUDA.compile), linker::typeof(CUDA.link))
    @ GPUCompiler C:\Users\Administrator\.julia\packages\GPUCompiler\OGnEB\src\execution.jl:262
  [6] cached_compilation(cache::Dict{…}, src::Core.MethodInstance, cfg::GPUCompiler.CompilerConfig{…}, compiler::Function, linker::Function)
    @ GPUCompiler C:\Users\Administrator\.julia\packages\GPUCompiler\OGnEB\src\execution.jl:151
  [7] macro expansion
    @ C:\Users\Administrator\.julia\packages\CUDA\sWPBr\src\compiler\execution.jl:373 [inlined]
  [8] macro expansion
    @ .\lock.jl:273 [inlined]
  [9] cufunction(f::GPUArrays.var"#gpu_broadcast_kernel_linear#38", tt::Type{…}; kwargs::@Kwargs{…})
    @ CUDA C:\Users\Administrator\.julia\packages\CUDA\sWPBr\src\compiler\execution.jl:368
 [10] macro expansion
    @ C:\Users\Administrator\.julia\packages\CUDA\sWPBr\src\compiler\execution.jl:112 [inlined]
 [11] (::KernelAbstractions.Kernel{…})(::CuArray{…}, ::Vararg{…}; ndrange::Tuple{…},
 workgroupsize::Nothing)
    @ CUDA.CUDAKernels C:\Users\Administrator\.julia\packages\CUDA\sWPBr\src\CUDAKernels.jl:103
 [12] _copyto!
    @ C:\Users\Administrator\.julia\packages\GPUArrays\uiVyU\src\host\broadcast.jl:71 [inlined]
 [13] materialize!
    @ C:\Users\Administrator\.julia\packages\GPUArrays\uiVyU\src\host\broadcast.jl:38 [inlined]
 [14] materialize!(dest::CuArray{…}, bc::Base.Broadcast.Broadcasted{…})
    @ Base.Broadcast .\broadcast.jl:880
 [15] top-level scope
    @ e:\Juliatest\cudatest.jl:9

caused by: CUDA error: no kernel image is available for execution on the device (code 209, ERROR_NO_BINARY_FOR_GPU)
Stacktrace:
  [1] checked_cuModuleLoadDataEx(_module::Base.RefValue{…}, image::Ptr{…}, numOptions::Int64, options::Vector{…}, optionValues::Vector{…})
    @ CUDA C:\Users\Administrator\.julia\packages\CUDA\sWPBr\lib\cudadrv\module.jl:28
  [2] CuModule(data::Vector{UInt8}, options::Dict{CUDA.CUjit_option_enum, Any})     
    @ CUDA C:\Users\Administrator\.julia\packages\CUDA\sWPBr\lib\cudadrv\module.jl:60
  [3] CuModule
    @ C:\Users\Administrator\.julia\packages\CUDA\sWPBr\lib\cudadrv\module.jl:49 [inlined]
  [4] link(job::GPUCompiler.CompilerJob, compiled::@NamedTuple{image::Vector{UInt8}, entry::String})
    @ CUDA C:\Users\Administrator\.julia\packages\CUDA\sWPBr\src\compiler\compilation.jl:414
  [5] actual_compilation(cache::Dict{…}, src::Core.MethodInstance, world::UInt64, cfg::GPUCompiler.CompilerConfig{…}, compiler::typeof(CUDA.compile), linker::typeof(CUDA.link))
    @ GPUCompiler C:\Users\Administrator\.julia\packages\GPUCompiler\OGnEB\src\execution.jl:262
  [6] cached_compilation(cache::Dict{…}, src::Core.MethodInstance, cfg::GPUCompiler.CompilerConfig{…}, compiler::Function, linker::Function)
    @ GPUCompiler C:\Users\Administrator\.julia\packages\GPUCompiler\OGnEB\src\execution.jl:151
  [7] macro expansion
    @ C:\Users\Administrator\.julia\packages\CUDA\sWPBr\src\compiler\execution.jl:373 [inlined]
  [8] macro expansion
    @ .\lock.jl:273 [inlined]
  [9] cufunction(f::GPUArrays.var"#gpu_broadcast_kernel_linear#38", tt::Type{…}; kwargs::@Kwargs{…})
    @ CUDA C:\Users\Administrator\.julia\packages\CUDA\sWPBr\src\compiler\execution.jl:368
 [10] macro expansion
    @ C:\Users\Administrator\.julia\packages\CUDA\sWPBr\src\compiler\execution.jl:112 [inlined]
 [11] (::KernelAbstractions.Kernel{…})(::CuArray{…}, ::Vararg{…}; ndrange::Tuple{…},
 workgroupsize::Nothing)
    @ CUDA.CUDAKernels C:\Users\Administrator\.julia\packages\CUDA\sWPBr\src\CUDAKernels.jl:103
 [12] _copyto!
    @ C:\Users\Administrator\.julia\packages\GPUArrays\uiVyU\src\host\broadcast.jl:71 [inlined]
 [13] materialize!
    @ C:\Users\Administrator\.julia\packages\GPUArrays\uiVyU\src\host\broadcast.jl:38 [inlined]
 [14] materialize!(dest::CuArray{…}, bc::Base.Broadcast.Broadcasted{…})
    @ Base.Broadcast .\broadcast.jl:880
 [15] top-level scope
    @ e:\Juliatest\cudatest.jl:9
Some type information was truncated. Use `show(err)` to see complete types.

Details on CUDA:

CUDA runtime 12.8, artifact installation
CUDA driver 12.8
NVIDIA driver 572.83.0

CUDA libraries:

CUBLAS: 12.8.3
CURAND: 10.3.9
CUFFT: 11.3.3
CUSOLVER: 11.7.2
CUSPARSE: 12.5.7
CUPTI: 2025.1.0 (API 26.0.0)
NVML: 12.0.0+572.83

Julia packages:

CUDA: 5.7.0
CUDA_Driver_jll: 0.12.0+0
CUDA_Runtime_jll: 0.16.0+0

Toolchain:

Julia: 1.11.4
LLVM: 16.0.6

The text was updated successfully, but these errors were encountered:

Yongda1 · 2025-03-24T10:54:42Z

sorry, I dont understand how to use Runic.jl to accept these changes on VS code. Could you show me how to do? Thanks.

Best regards
Jim

maleadt · 2025-03-24T11:13:36Z

how to use Runic.jl to accept these changes on VS code

There's nothing you need to do. Runic.jl is a code formatter, and that comment is intended for the PR author (i.e., me).

Yongda1 · 2025-03-24T11:15:38Z

Thank you.

Yongda1 · 2025-03-24T12:35:38Z

Excuse me, so, I just need wait for the next version?

maleadt · 2025-03-24T12:50:10Z

You can check out the master branch of CUDA.jl for the time being.

Yongda1 · 2025-03-24T12:55:46Z

Thanks. Could you tell me how to install the master branch on VS Code? I only know Pkg.add("CUDA").

maleadt · 2025-03-24T13:28:58Z

https://pkgdocs.julialang.org/v1/managing-packages/#Adding-packages

If a branch (or a certain commit) of Example has a hotfix that is not yet included in a registered version, we can explicitly track that branch (or commit) by appending #branchname (or #commitSHA1) to the package name

Yongda1 · 2025-03-24T13:40:34Z

Thanks.
I used this line,
Pkg.add(url="https://github.com/JuliaGPU/CUDA.jl")
However, when i am running the following codes, I still got error.

import Pkg
using Test
Pkg.add(url="https://github.com/JuliaGPU/CUDA.jl")
#CUDA.versioninfo()
N = 100
x_d = CUDA.fill(1.0f0, N)  # a vector stored on the GPU filled with 1.0 (Float32)
y_d = CUDA.fill(2.0f0, N)  # a vector stored on the GPU filled with 2.0
y_d .+= x_d
@test all(Array(y_d) .== 3.0f0)

ERROR: 
Stacktrace:
  [1] error(s::String)
    @ Base .\error.jl:35
  [2] CuModule(data::Vector{UInt8}, options::Dict{CUDA.CUjit_option_enum, Any})
    @ CUDA C:\Users\Administrator\.julia\packages\CUDA\sWPBr\lib\cudadrv\module.jl:69
  [3] CuModule
    @ C:\Users\Administrator\.julia\packages\CUDA\sWPBr\lib\cudadrv\module.jl:49 [inlined]
  [4] link(job::GPUCompiler.CompilerJob, compiled::@NamedTuple{image::Vector{UInt8}, entry::String})
    @ CUDA C:\Users\Administrator\.julia\packages\CUDA\sWPBr\src\compiler\compilation.jl:414
  [5] actual_compilation(cache::Dict{…}, src::Core.MethodInstance, world::UInt64, cfg::GPUCompiler.CompilerConfig{…}, compiler::typeof(CUDA.compile), linker::typeof(CUDA.link))
    @ GPUCompiler C:\Users\Administrator\.julia\packages\GPUCompiler\OGnEB\src\execution.jl:262
  [6] cached_compilation(cache::Dict{…}, src::Core.MethodInstance, cfg::GPUCompiler.CompilerConfig{…}, compiler::Function, linker::Function)
    @ GPUCompiler C:\Users\Administrator\.julia\packages\GPUCompiler\OGnEB\src\execution.jl:151
  [7] macro expansion
    @ C:\Users\Administrator\.julia\packages\CUDA\sWPBr\src\compiler\execution.jl:373 [inlined]
  [8] macro expansion
    @ .\lock.jl:273 [inlined]
  [9] cufunction(f::GPUArrays.var"#gpu_broadcast_kernel_linear#38", tt::Type{…}; kwargs::@Kwargs{…})
    @ CUDA C:\Users\Administrator\.julia\packages\CUDA\sWPBr\src\compiler\execution.jl:368
 [10] macro expansion
    @ C:\Users\Administrator\.julia\packages\CUDA\sWPBr\src\compiler\execution.jl:112 [inlined]
 [11] (::KernelAbstractions.Kernel{…})(::CuArray{…}, ::Vararg{…}; ndrange::Tuple{…}, workgroupsize::Nothing)
    @ CUDA.CUDAKernels C:\Users\Administrator\.julia\packages\CUDA\sWPBr\src\CUDAKernels.jl:103
 [12] _copyto!
    @ C:\Users\Administrator\.julia\packages\GPUArrays\uiVyU\src\host\broadcast.jl:71 [inlined]
 [13] materialize!
    @ C:\Users\Administrator\.julia\packages\GPUArrays\uiVyU\src\host\broadcast.jl:38 [inlined]
 [14] materialize!(dest::CuArray{…}, bc::Base.Broadcast.Broadcasted{…})
    @ Base.Broadcast .\broadcast.jl:880
 [15] top-level scope
    @ e:\Juliatest\cudatest.jl:8

caused by: CUDA error: no kernel image is available for execution on the device (code 209, ERROR_NO_BINARY_FOR_GPU)
Stacktrace:
  [1] checked_cuModuleLoadDataEx(_module::Base.RefValue{…}, image::Ptr{…}, numOptions::Int64, options::Vector{…}, optionValues::Vector{…})
    @ CUDA C:\Users\Administrator\.julia\packages\CUDA\sWPBr\lib\cudadrv\module.jl:28
  [2] CuModule(data::Vector{UInt8}, options::Dict{CUDA.CUjit_option_enum, Any})
    @ CUDA C:\Users\Administrator\.julia\packages\CUDA\sWPBr\lib\cudadrv\module.jl:60
  [3] CuModule
    @ C:\Users\Administrator\.julia\packages\CUDA\sWPBr\lib\cudadrv\module.jl:49 [inlined]
  [4] link(job::GPUCompiler.CompilerJob, compiled::@NamedTuple{image::Vector{UInt8}, entry::String})
    @ CUDA C:\Users\Administrator\.julia\packages\CUDA\sWPBr\src\compiler\compilation.jl:414
  [5] actual_compilation(cache::Dict{…}, src::Core.MethodInstance, world::UInt64, cfg::GPUCompiler.CompilerConfig{…}, compiler::typeof(CUDA.compile), linker::typeof(CUDA.link))
    @ GPUCompiler C:\Users\Administrator\.julia\packages\GPUCompiler\OGnEB\src\execution.jl:262
  [6] cached_compilation(cache::Dict{…}, src::Core.MethodInstance, cfg::GPUCompiler.CompilerConfig{…}, compiler::Function, linker::Function)
    @ GPUCompiler C:\Users\Administrator\.julia\packages\GPUCompiler\OGnEB\src\execution.jl:151
  [7] macro expansion
    @ C:\Users\Administrator\.julia\packages\CUDA\sWPBr\src\compiler\execution.jl:373 [inlined]
  [8] macro expansion
    @ .\lock.jl:273 [inlined]
  [9] cufunction(f::GPUArrays.var"#gpu_broadcast_kernel_linear#38", tt::Type{…}; kwargs::@Kwargs{…})
    @ CUDA C:\Users\Administrator\.julia\packages\CUDA\sWPBr\src\compiler\execution.jl:368
 [10] macro expansion
    @ C:\Users\Administrator\.julia\packages\CUDA\sWPBr\src\compiler\execution.jl:112 [inlined]
 [11] (::KernelAbstractions.Kernel{…})(::CuArray{…}, ::Vararg{…}; ndrange::Tuple{…}, workgroupsize::Nothing)
    @ CUDA.CUDAKernels C:\Users\Administrator\.julia\packages\CUDA\sWPBr\src\CUDAKernels.jl:103
 [12] _copyto!
    @ C:\Users\Administrator\.julia\packages\GPUArrays\uiVyU\src\host\broadcast.jl:71 [inlined]
 [13] materialize!
    @ C:\Users\Administrator\.julia\packages\GPUArrays\uiVyU\src\host\broadcast.jl:38 [inlined]
 [14] materialize!(dest::CuArray{…}, bc::Base.Broadcast.Broadcasted{…})
    @ Base.Broadcast .\broadcast.jl:880
 [15] top-level scope
    @ e:\Juliatest\cudatest.jl:8
Some type information was truncated. Use `show(err)` to see complete types.

maleadt · 2025-03-24T13:45:04Z

Sorry, I hadn't looked closely enough at the error message; this isn't simply a compatibility issue.

maleadt · 2025-03-25T13:38:54Z

I'm afraid this may have to wait until I get a hold on Blackwell hardware, which hopefully happens in a couple of weeks. In the mean time, if somebody's interested feel free to take a look. We're probably invoking nvcc incorrectly, e.g., targeting a wrong compute capability or PTX ISA.

Yongda1 · 2025-03-26T04:49:27Z

It is ok. Thanks for your effort.

zipeilee · 2025-04-04T18:27:31Z

I took a look at merge #2717 by @maleadt , and it seems that LLVM requires v20 or higher. Could it be that the LLVM version is too low? Currently, my Julia 1.11.4 comes with LLVM v16. I’m not sure if this information is helpful to you, or if there is a way to resolve this issue.

Additionally, I can provide my cuda.versioninfo() output for my 5070 Ti if needed.

CUDA runtime 12.8, artifact installation
CUDA driver 12.8
NVIDIA driver 570.133.7

CUDA libraries: 
- CUBLAS: 12.8.4
- CURAND: 10.3.9
- CUFFT: 11.3.3
- CUSOLVER: 11.7.3
- CUSPARSE: 12.5.8
- CUPTI: 2025.1.1 (API 26.0.0)
- NVML: 12.0.0+570.133.7

Julia packages: 
- CUDA: 5.7.1
- CUDA_Driver_jll: 0.12.1+1
- CUDA_Runtime_jll: 0.16.1+0

Toolchain:
- Julia: 1.11.4
- LLVM: 16.0.6

2 devices:
  0: NVIDIA GeForce RTX 5070 Ti (sm_120, 9.878 GiB / 15.921 GiB available)
  1: NVIDIA GeForce RTX 3070 Ti (sm_86, 6.953 GiB / 8.000 GiB available)

maleadt · 2025-04-09T08:29:17Z

It seems that LLVM requires v20 or higher.

Not necessarily; we invoke ptxas manually using a higher .version attribute:

CUDA.jl/src/compiler/compilation.jl

Lines 145 to 149 in eb4fcad

    
           # if LLVM couldn't target the requested PTX ISA, bump it in the assembly. 
        
           if job.config.target.ptx != job.config.params.ptx 
        
               ptx = job.config.params.ptx 
        
               asm = replace(asm, r"(\.version .+)" => ".version $(ptx.major).$(ptx.minor)") 
        
           end

Yongda1 · 2025-04-09T08:41:12Z

Thanks. I will try it later.

Yongda1 added the bug Something isn't working label Mar 23, 2025

maleadt mentioned this issue Mar 24, 2025

Add compat entries for recent devices and toolkits. #2717

Merged

maleadt closed this as completed in #2717 Mar 24, 2025

maleadt reopened this Mar 24, 2025

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

RTX5070Ti is not fully supported by CUDA 12.8.0 #2714

RTX5070Ti is not fully supported by CUDA 12.8.0 #2714

Yongda1 commented Mar 23, 2025

Yongda1 commented Mar 24, 2025

maleadt commented Mar 24, 2025

Yongda1 commented Mar 24, 2025

Yongda1 commented Mar 24, 2025

maleadt commented Mar 24, 2025

Yongda1 commented Mar 24, 2025

maleadt commented Mar 24, 2025

Yongda1 commented Mar 24, 2025

maleadt commented Mar 24, 2025

maleadt commented Mar 25, 2025

Yongda1 commented Mar 26, 2025

zipeilee commented Apr 4, 2025

maleadt commented Apr 9, 2025

Yongda1 commented Apr 9, 2025

RTX5070Ti is not fully supported by CUDA 12.8.0 #2714

RTX5070Ti is not fully supported by CUDA 12.8.0 #2714

Comments

Yongda1 commented Mar 23, 2025

Yongda1 commented Mar 24, 2025

maleadt commented Mar 24, 2025

Yongda1 commented Mar 24, 2025

Yongda1 commented Mar 24, 2025

maleadt commented Mar 24, 2025

Yongda1 commented Mar 24, 2025

maleadt commented Mar 24, 2025

Yongda1 commented Mar 24, 2025

maleadt commented Mar 24, 2025

maleadt commented Mar 25, 2025

Yongda1 commented Mar 26, 2025

zipeilee commented Apr 4, 2025

maleadt commented Apr 9, 2025

Yongda1 commented Apr 9, 2025