Amber24 GPU run fails on RTX 5090 – “no kernel image is available for execution on the device”

Accelerated Computing CUDA CUDA Programming and Performance
1

Hello everyone,

I am facing an issue while running molecular dynamics simulations using Amber24 with GPU acceleration.

System details:

  • GPU: NVIDIA GeForce RTX 5090

  • CUDA Toolkit: 12.4

  • Driver: NVIDIA-SMI 570.211.01

  • OS: Ubuntu 22.04

  • Compiler: GCC 11.4

Build details:

  • Amber24 compiled successfully with CUDA support (pmemd.cuda builds without errors)

  • Installation completed without any issues

Problem:

When I try to run an MD simulation using pmemd.cuda, the job fails immediately with the following error:

cudaMemcpyToSymbol: SetSim copy to cSim failed no kernel image is available for execution on the device

Additional observations:

  • GPU is correctly detected (nvidia-smi shows the RTX 5090)

  • pmemd.cuda starts and prints GPU device info, but crashes before simulation begins

  • I have already tried:

    • Recompiling Amber multiple times

    • Using CUDA 12.4 (also tested with newer toolkit earlier)

    • Setting CUDA_VISIBLE_DEVICES

    • Cleaning and rebuilding from scratch

Request:

Could someone please guide:

  • How to ensure proper GPU architecture support during compilation?

  • Whether CUDA 12.4 fully supports RTX 5090 in this context?

  • Any known compatibility issues between Amber24 and newer GPUs?

Any help or suggestions would be greatly appreciated.

Thank you!

2

Hi! I have resolved the error with the next configuration.

GPU: NVIDIA GeForce RTX 5060ti

CUDA Toolkit: 12.8

Driver: NVIDIA-SMI 595.71.05

OS: Ubuntu 24.04.4 LTS

Compiler: GCC 10.5.0

The test give some minor errors but pmemd.cuda run well.

Good luck