Kernel error: forward compatibility was attempted on non supported HW

Accelerated Computing CUDA CUDA Setup and Installation
1

Hi Experts,

I previously ran the project smoothly on the 4060 Ada hardware, but when I moved the exact same project to the 5090 Blackwell, this error appeared.

I’m providing my version details below. Could you please help me identify if there is any misconfiguration? Thank you.

cmake:

cmake_minimum_required(VERSION 3.14)
project(test LANGUAGES CXX CUDA)

set(CMAKE_CUDA_ARCHITECTURES 75 86 87 89 90 90a CACHE STRING "CUDA architectures")
set(CMAKE_CXX_STANDARD 17)
set(CMAKE_CXX_STANDARD_REQUIRED ON)
set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS} --default-stream per-thread -std=c++17 --expt-relaxed-constexpr -Xcompiler=-fPIE -Xcudafe --diag_suppress=611")
set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -Xcompiler=-fopenmp")
set(CMAKE_CUDA_FLAGS_RELEASE "-O3 -lineinfo -std=c++17 --extended-lambda -Xcudafe --diag_suppress=611")
set(CMAKE_CUDA_FLAGS_DEBUG "-O0 -std=c++17 --extended-lambda -Xcudafe --diag_suppress=611 -g -G")

NVIDIA-SMI 570.86.10 
Driver Version: 570.86.10 
CUDA Version: 12.8
2

You probably want to add 120 to the line:

set(CMAKE_CUDA_ARCHITECTURES 75 86 87 89 90 90a CACHE STRING "CUDA architectures")

You need Cuda Toolkit 12.8 or greater.

3

What is sm120? Blackwell is 90, right?

4

No, Blackwell (the 50XX variant), is 120. Hopper is 90.

5 
cmake_minimum_required(VERSION 3.14)
project(test LANGUAGES CXX CUDA)

set(CMAKE_CUDA_ARCHITECTURES 75 86 87 89 90 120 CACHE STRING "CUDA architectures")
set(CMAKE_CXX_STANDARD 17)
set(CMAKE_CXX_STANDARD_REQUIRED ON)
set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS} --default-stream per-thread -std=c++17 --expt-relaxed-constexpr -Xcompiler=-fPIE -Xcudafe --diag_suppress=611")
set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -Xcompiler=-fopenmp")
set(CMAKE_CUDA_FLAGS_RELEASE "-O3 -lineinfo -std=c++17 --extended-lambda -Xcudafe --diag_suppress=611")
set(CMAKE_CUDA_FLAGS_DEBUG "-O0 -std=c++17 --extended-lambda -Xcudafe --diag_suppress=611 -g -G")

Still see this error: forward compatibility was attempted on non supported HW

What else I can check?

I’m using this docker image
nvidia/cuda:12.8.0-cudnn-devel-ubuntu22.04

圖片
圖片742×326 17.7 KB

6

Sorry, I have nothing else to offer.